• Report claims an incredible 80% of AI tools are running in organizations without IT oversight
• Browser agents and integration tools are escaping the attention of security and IT engineers
• Reco’s State of Agent Security 2026 report also tracked 637 vulnerabilities across agents and LLMs

Unmonitored deployment of AI tools is a risk to security, and puts data at risk, a new study has claimed.

The report from Reco, which draws its information from disclosed vulnerabilities, analysis of 500 Model Context Protocol servers, and Reco’s own platform telemetry, found four in five AI tools (80%) are running without oversight from IT departments.

Of particular concern is the scale of AI applications in use. Smaller companies use 414 AI tools per 1,000 employees without IT approval, apparently a combination of browser extensions and workflows beyond the usual review and approval process.

Data risks from unmonitored AI

Giving AI tools to employees might unlock productivity boosts, but their use has to be approved. That’s the key takeaway from the report, which highlights some concerning cybersecurity figures. For example, it assessed 500 agent tools and found “62% can both read local data and reach the internet.” This represents an opportunity for data exfiltration.

Elsewhere, 637 AI agent related vulnerabilities were identified in the report.

The adoption of AI is wider than specific use of a SaaS application or visiting ChatGPT. Reco found that AI agents are running within other tools, and inheriting user permissions. The implications of this are clear.

Operational risk

Analysis of the telemetry (gathered from 62 enterprise-scale businesses in financial services, healthcare, retail, and telecommunications between January 1 and August 1 2026) reveals a free-for-all attitude towards AI adoption. While businesses may have policies and procedures in place and processes to assess, evaluate, review, and finally approve new AI-based tools, these are being circumvented for low-level applications.

The rules work for SaaS procurement oversight, but not for browser extensions, and the result is “operational risk.”

“AI agents have moved from experimentation into daily business workflows, but our findings show only 20% of AI tools in enterprise ecosystems are currently governed by IT oversight," Reco CEO Ofer Klein noted.

“That leaves organizations exposed to a new class of operational risk. Agents embedded in applications can operate through existing permissions, OAuth grants and workflow access, creating toxic combinations that expose data and trigger actions beyond what any owner approved.”

Organizations will need to give IT teams the resources they need to manage and restrict unauthorized AI use, as the alternative means leaving the gates open to the possibility of data exfiltration.

• GMKtec NucBox M5 Ultra pairs an eight-core Ryzen 7 7730U with dual 2.5GbE, Wi-Fi 6E, and triple-display output while starting at just $330 a pop for a barebones configuration
• The underlying 2023-release CPU is capable, but a rebadge of the 2022-launched Zen 3 Ryzen 7 5825U, which is, in turn, a 2021-launched 5800U with a 100 MHz bump to base and boost clocks
• The price can be reduced to as little as $280 per unit for users buying 5 of the RAM- and SSD-less SKU

GMKtec has put its NucBox M5 Ultra back in the limelight, and the offering is familiar for the Shenzhen mini PC maker's mid-range lineup.

It comes with an eight-core AMD processor, two 2.5 Gigabit Ethernet ports, upgradeable memory and storage, and a price starting just over $300.

The catch for any prospective buyer is that you must bring your own DDR4 laptop RAM and SSD to get it going.

A capable NUC that is hiding its underlying CPU's age

The M5 Ultra runs the AMD Ryzen 7 7730U, an eight-core, sixteen-thread part clocked from 2.0 GHz to 4.5 GHz with 16 MB of L3 cache, built on TSMC's 7nm process with a default 15W TDP. Despite the 7000-series badge, it belongs to AMD's Cezanne lineup via the Barcelo-U refresh announced in January 2022.

The CPU it most closely matches in clock speeds, core configuration, and performance is the Ryzen 7 5825U, which is a slightly higher-clocked version of the January 2021-released Ryzen 7 5800U, which runs 100MHz lower on both base and boost clocks.

This is interesting because prospective buyers might feel they are buying a Zen 4-based CPU thanks to its 7000-series badge but would end up getting a CPU that was first available more than half a decade ago, even if it does hold its own in its price and performance segment.

A well-rounded offering with plenty of expansion under the hood

The M5 Ultra offers two DDR4-3200 SODIMM slots supporting up to 64 GB, and two M.2 2280 slots, allowing for a variety of configurations based on the user's needs.

It does have its limitations due to the underlying CPU, however: while GMKtec claims each M2 slot can support up to 8 TB, for a nominal 16 TB ceiling, speeds are capped at PCIe 3.0 because the 7730U provides only 16 usable lanes and no Gen 4 support out of the box.

Connectivity is where the unit particularly shines: it comes with two Realtek RTL8125 2.5GbE controllers, Wi-Fi 6E and Bluetooth 5.2, and a 10 Gbps, 100W USB port in addition to two USB 3.2 Gen 2 Type-A ports and two USB 2.0 ports, a very formidable offering for a Nuc-tier PC.

The system, which is powered by a 65W PSU over a Type-C connector, also offers HDMI 2.0, DisplayPort, and USB-C for up to three simultaneous connections for power users looking to run a multi-display setup.

One could argue that pricey RAM and SSD options make sourcing them considerably less appealing and a tall order, but even with that in mind, at today's prices, the NucBox has plenty of connectivity under the hood, decent power, and a raw customizability that is hard to argue with in 2026.

• Think tank Ceres found power generation for data centers in seven states withdraws 3.4 trillion gallons of freshwater a year, or 10.4 million acre-feet
• The figure might be a bit overstated, as it counts withdrawals, including hydropower water returned to the river immediately
• 78% of electricity in those states comes from water-using plants, and 66% of those plants sit in water-stressed areas

Non-profit think tank Ceres has published a new report analyzing the water used to generate electricity for data centers across Virginia, Texas, California, Illinois, Georgia, Ohio, and Arizona.

Ceres estimates data centers use 3.4 trillion gallons of freshwater a year, or 10.4 million acre-feet.

Bloomberg Law's Bobby Magill compared that unfavorably against the 8 million acre-feet California uses annually for urban purposes, a figure from a 2021 Legislative Analyst's Office report.

An important difference withdrawal versus consumption

Ceres's approach makes sense, but it paints only a partial picture at best. It measures water withdrawals, or the total volume of water a power company pulls from a water source like a river, lake, or aquifer, including water it essentially replaces shortly afterward.

For context, the LAO's 8 million acre-feet is a different quantity: water actually delivered and applied to residential, commercial, industrial, and landscape uses across California, and it is unlikely to be channeled back to its source.

Ceres does note this, stating that "withdrawal reflects the full volume an asset depends on to operate", which it presents as a strong indicator of risk exposure to multiple communities that are already living in drought-affected areas.

A large amount of the power being generated and measured is hydroelectric power. This makes it a bit ironic, given that using water for generation may not be considered 'consumed' effectively.

The number seems considerably higher than one would expect: Arizona State University researchers, using 2024 Lawrence Berkeley National Laboratory data, put the indirect consumption of US data centers, including power generation, at about 650,000 acre-feet a year nationwide.

Although the data are old, the numbers suggest data centers may be consuming a fraction of what Ceres implies, according to its own report, unless consumption rose drastically to 16 times its reported figure.

Ceres, however, for now projects that water for power generation could account for 72 percent of total US data center water consumption by 2030, on a consumption basis, even as cooling inside the buildings gets more efficient.

At a time when the two largest Colorado River reservoirs have hit record lows, Ceres noted that local opposition grounded in water and grid concerns disrupted $130 billion of data center projects in the first quarter of 2026 alone.

Shama Perveen, the report's co-author and Ceres' director of water research, framed this as a complex situation and one that required a more inclusive approach, saying "no single stakeholder can address these challenges alone".

Whether it pans out at a time when AI companies are spending money hand over fist to get their hyperscale operations (and subsequent cooling and power requirements) out of the door as the search for AGI continues unabated remains to be seen, however.

• Intel report claims 60% of business leaders expect to operate robot fleets within the next five years
• Conversely, only 40% have a human-robot workforce strategy
• The Robotics Readiness Gap report assesses the state of robotic adoption around the world

Intel’s 'Robotics Readiness Gap' report has revealed that many large businesses are planning to implement a human-robot workforce within the next five years, with a majority expecting a doubling of operational output with a full-scale deployment.

The survey of 800 senior leaders in organizations with over $500 million annual turnover also found while adoption plans are accelerating, the rollout of a human-robot workforce has some key challenges, not least infrastructure, skills, strategy, and safety.

Despite the positive outlook for robotics, the scenario has a heavy dose of reality. Only 40% of businesses surveyed have a suitable strategy in place, despite a belief that robots could prove more cost-effective than humans within three years.

Human-robot workforce

Within this context, the aim of achieving a human-robot workforce seems more achievable, although some of the findings suggest an aspirational element to the drive towards robotics.

While 60% of leaders are expecting to run a fleet of robots by 2030, only 40% have a suitable human-robot workforce strategy in place. So while 74% believe workforce planning will be inseparable from robotics strategy within 5 years, there is a gap in preparation.

The report notes that “On average, leaders believe that deploying a robot will become more cost-effective than hiring a human in their sector within three years.”

But does this mean replacing jobs? Intel says not. “This isn’t about replacing people but creating new forms of human-machine collaboration, with robots working alongside employees as productive teammates.”

Doubling operational output

Intel’s report states that robotics in the workplace is approaching a “critical threshold,” with leaders predicting that “operational output would double if robotics were deployed at full scale across their organization.”

However, both notions rely heavily on five factors: strategy, skills, safety, shape (function), and scale (which includes physical AI infrastructure).

Intel’s report states, “few are prepared for what large-scale deployment actually requires. Closing this gap will mean redesigning operating models for a workforce that includes autonomous systems.” Even in the defense industry, this gap exists (“confidence outstrips preparedness by 48%”).

Placed in these terms, and with the context of the size of the organizations quizzed on their robotics strategies, that critical threshold seems a little further away than five years.

• AMD Ryzen AI Max+ PRO 495 mini PCs now reach 192 GB of memory
• GMKtec says its EVO X5 Pro can run 300B parameter models locally
• Framework Desktop combines 16 cores with 192GB unified memory while Acemagic fits the PRO 495 inside a 2L case

After AMD launched the Ryzen AI Max+ PRO 495 in May 2026, the processor is now appearing in a new wave of mini PCs offering up to 192GB of memory.

The chip follows 2025’s Ryzen AI MAX+ 395, which appeared in at least 37 mini PCs costing above $1,500.

Acemagic, GMKtec and Framework have now shown systems using the newer processor, with GMKtec claiming support for 300B models.

New mini PCs push memory far beyond ordinary systems

Framework is preparing a Desktop model featuring the PRO 495 alongside 192GB of LPDDR5X memory and 273GB/s bandwidth, combining 16 Zen 5 cores, 32 threads, 40 graphics compute units, and 131 TOPS for AI workloads.

The integrated Radeon 8065S graphics shares system memory, allowing applications to access resources without relying on separate graphics memory.

Framework says the machine will occupy 4.5 litres, while Linux can also come preloaded for users who choose that option.

GMKtec has unveiled its EVO X5 Pro at IFA 2026, using the same processor with up to 192GB memory.

The system uses LPDDR5X running at 8,533 MT/s, while memory bandwidth reaches approximately 273GB/s under the company’s stated specifications.

GMKtec says users can assign as much as 160GB of that memory to graphics processing when required.

According to GMKtec, the processor reaches 5.2GHz, while its Radeon graphics include 40 compute units and hardware ray tracing support.

The company says this configuration can run large language models containing as many as 300B parameters without requiring cloud services.

GMKtec also includes an XDNA 2 neural processor capable of delivering up to 55 TOPS for supported artificial intelligence workloads.

The EVO X5 Pro has four USB4 connections, external graphics support, 8K output, and networking options for linked systems.

A vapor chamber, three fans, and three operating modes are used to manage heat during sustained computing workloads.

Acemagic adds another compact option with 192GB

Acemagic’s F9A uses the PRO 495 with as much as 192GB of LPDDR5X memory inside a roughly 2-litre enclosure.

The computer measures 158.5 × 158.5 × 81.5 mm, making its physical size considerably smaller than Framework’s 4.5-litre machine.

Its processor reaches 5.1GHz, while the Radeon 8065S graphics unit operates at frequencies reaching 3.0GHz under suitable conditions.

The system provides two PCIe 4.0 NVMe slots, allowing storage configurations using RAID 0 or RAID 1 when supported.

Acemagic also includes OCuLink using PCIe 4.0 x4, enabling compatible external graphics cards and other high-speed peripherals.

Two USB4 Type-C ports provide transfer rates reaching 40Gbps, while networking includes Wi Fi 7 and dual 2.5GbE.

The company says the F9A can run DeepSeek V4 Flash locally, with that model containing 284B parameters.

As of the time of writing, none of the three brands has published final retail pricing for its respective configurations.

Given the hardware involved and the premium attached to the previous generation, prices above $2,000 would not be surprising.

• Microsoft faces claims that its data center noise harms nearby residents
• Standard dBA monitors may fail to capture some low-frequency noise
• Residents accuse xAI of causing noise, vibration, and air pollution

Microsoft and xAI are facing lawsuits over noise and other alleged harms linked to data center operations in Wisconsin and Mississippi.

At least six similar cases have been filed this year, as residents challenge the impacts from currently active facilities.

In addition to nuisance, the disputes also focus on negligence claims, with plaintiffs seeking damages or changes to data center operations.

Low-frequency noise creates a measurement problem

Microsoft faces a proposed class action concerning its Fairwater facility in Mount Pleasant, Wisconsin, which began operating in June 2026.

Residents filed the case in late July, alleging that persistent low-frequency noise affects their homes and property values, and arguing standard dBA equipment may fail to detect the sound because its frequency falls outside effective measurement ranges.

“dBA measures do not effectively detect lower frequency sounds like those emitted by Defendant’s Data Center operations,” the lawsuit says.

“‘Data center cooling — massive chillers, slow turning fans, compressor arrays — generates exactly the long wavelength noise’” that is not adequately measured using dBA.”

The plaintiffs argue Microsoft could have reduced the disturbance through acoustic barriers, quieter cooling equipment, soundproofing, and improved monitoring.

“A properly operated, maintained, and/or constructed Data Center will contain, capture, or otherwise prevent the emission of excessive noise from its generators and cooling systems,” the filing says.

That argument could create difficulties for facilities where regulatory measurements fail to capture sounds that residents can still hear.

The dispute also raises questions about how noise from data centers should be measured.

xAI faces claims involving turbines and pollution

In Mississippi, residents have sued xAI over alleged noise, vibration, and air pollution from dozens of gas turbines.

The turbines supply electricity for the company's Southaven data center, according to allegations contained in the lawsuit.

Attorneys representing residents accused xAI of failing to take meaningful action to address the alleged effects.

“xAI has made no meaningful effort to halt the harm,” attorneys at Weitz & Luxenberg said in a statement.

Similar lawsuits have also been filed in Vineland, New Jersey, Dowagiac, Michigan, North Tonawanda, New York, and Hood County, Texas.

The cases generally rely on nuisance or negligence claims concerning conditions plaintiffs believe operators could have prevented or controlled.

Several other data center disputes remain unresolved, with none of the noise cases reaching summary judgment or trial.

However, other environmental nuisance cases have already produced settlements involving substantial sums of money.

Amazon agreed to pay more than $20 million over nitrate contamination linked to one Oregon data center.

The company said officials settled early to avoid prolonged litigation and concentrate resources on community support instead.

Microsoft had not immediately responded to a request for comment concerning the Wisconsin lawsuit when the report was published.

Via SmartCitiesDrive

• Huawei MateBook Pro S weighs 798g at 11.9mm, with a 14.2-inch 3.1K 120Hz flexible OLED touchscreen and a 54Wh battery
• The laptop runs HarmonyOS 6.1 rather than Windows and uses a Kirin XE90 SoC, a HiSilicon ARM design on SMIC's 7nm node, capped at 20W
• It currently sells only in China, where it starts at 7,999 yuan, with no plans announced for wider availability

Huawei has launched the MateBook Pro S, a 14.2-inch laptop that weighs 798 grams and measures 11.9mm at its thickest point.

The device runs HarmonyOS 6.1 rather than Windows, uses Huawei's custom Kirin XE90 processor, and went on sale in China on 14 August, with prices starting at 7,999 yuan, roughly just under $1,200 when converted.

There seem to be no documented plans to offer the same laptop officially in the West, but third-party suppliers have already emerged offering to ship to countries worldwide, albeit at a significant premium.

Thin, but not the thinnest, with a CPU that is relatively untested

At 798 grams, the MateBook Pro S is noticeably lighter than most of its competition, but thinner alternatives from Japanese manufacturers exist, shaving off an additional 164 grams.

Although it isn't one of the lightest in the business, it is considerably slimmer than laptops that compete in the same performance segments. Huawei achieves this with a magnesium-lithium alloy bottom cover, magnesium-aluminum A and C covers, a three-section motherboard, a thin dual-fan cooling assembly with what Huawei calls Shark Fin blades, and a high-silicon-anode battery.

With a 3.1k flexible OLED touchscreen offering 1000 nits of full-screen brightness and a peak of 1600 nits, plus a 120Hz refresh rate, it has one of the best displays on paper.

What stands out even more than the capable display, however, is the processor: the Kirin XE90, as Huawei claims 23% higher single-core performance, 25 percent better energy efficiency, and 40% higher NPU throughput than the standard Kirin X90, with the cooling system supporting up to 20W of sustained output.

These numbers, however, are unchecked, and much of what one can find about the laptop's performance is inferred from comparisons using Huawei's own figures or paid advertisements claiming it delivers 'formidable power'.

More interestingly, it may use the same SMIC 7nm N+2 node TechInsights found when it physically analyzed the Kirin X90 in the MateBook Fold, rather than the 5nm process Chinese television claimed. The XE90 essentially is a refinement of the same design, and there is no indication it has moved nodes since the finding was reported.

This matters because it affects not only Huawei's laptop performance, but also efficiency in a class where performance per watt is a key determinant. The laptop is also only 798 grams in its lightest configuration; its lineup splits into Standard, Soft Light, Privacy, and Collector's Edition, and the weights differ considerably. The standard and privacy models weigh 798g, the Soft Light version 805g, and the Collector's Edition 825g, with a thickness of 12.2mm.

Huawei also claims 18 hours of video playback, a stable WiFi 7+ connection up to 1000 meters away in open, unobstructed areas, and a six-microphone array that can clearly capture audio from up to 10 meters away.

For those looking for a solution that isn't Windows- or Linux-based, the MateBook Pro S is an intriguing buy — if you're willing to pay the premium.

• REGENT Viceroy Seaglider has completed its first human-crewed flight above Narragansett Bay
• The electric craft flew 1,956 feet during its first crewed test
• Viceroy can shift between floating, hydrofoiling, and low-altitude flight modes

REGENT Craft has completed the first crewed flight of its Viceroy Seaglider, advancing development of an electric vessel designed for coastal transport.

The prototype carried two captains during a 30-second test above Narragansett Bay, reaching 33 ft while travelling 1,956 ft.

REGENT says the flight marks the largest human-carrying all-electric wing-in-ground-effect craft to complete such a test.

First crewed flight tests an unusual transport concept

Viceroy combines a boat-like hull, underwater foils, and wings, allowing it to shift between water travel and low-altitude flight.

At low speeds, the craft remains afloat before hydrofoils lift its hull and prepare it for flight above the surface.

During test flight, the vehicle remains close to the water, operating within one wingspan of the surface rather than climbing conventionally.

"Completing this flight test validates years of rigorous engineering and operational work and gives us the confidence to move into full-rate production,” said Mike Klinker, co-founder and CTO at REGENT.

“We showed today that flying a Seaglider is every bit as smooth and safe as we designed it to be.”

REGENT says the production model will carry 12 passengers and two crew, with a maximum takeoff weight of 15,000 lbs.

Its planned flight speed reaches 180 mph, while its stated range extends to 180 miles between coastal destinations.

The company says passengers will have large windows, substantial legroom, and lower cabin noise than conventional aircraft during travel.

The craft uses a three-layer digital control system intended to manage stability while allowing mariners to operate it like watercraft.

REGENT says certified mariners will captain the vessels, while United States Coast Guard rules will govern operations within American waters.

Commercial plans expand alongside defence work

These vessels are designed for coastal routes, where existing ports could support operations without conventional airport infrastructure.

About 40% of the global population lives in coastal communities, providing a large potential market for such services.

At the moment, the company has booked orders spanning six continents, while manufacturing capacity has been reserved across several future production years.

Earlier this year, REGENT announced an agreement involving The Twenty Five, which ordered as many as 60 Viceroy vessels.

The arrangement would make the operator REGENT's first United States delivery and launch partner for commercial Seaglider services.

Military versions for logistics, medical evacuation, surveillance missions, and other maritime applications requiring rapid movement are in development

The company holds a $15 million agreement with the United States Marine Corps Warfighting Laboratory for related development work.

REGENT completed construction of a 255,000-square-foot manufacturing facility in Rhode Island during June, with production activities beginning this year.

The facility is intended for structural assembly, wing and foil fitting, battery installation, systems work, and water-based testing.

“This is the moment we have been envisioning and tirelessly pursuing for the past five and a half years: REGENT has successfully proven safe, efficient ground-effect flight,” said Billy Thalheimer, co-founder and CEO at REGENT.

• GPUThor is the first Rowhammer attack to defeat ECC on Nvidia GPUs and reach a root shell on the host
• It affects four Ampere workstation cards with GDDR6 memory, the RTX A4000, A4500, A5000 and A6000
• Non-uniform hammering, enabled by defeating GPU memory coalescing and finding that Target Row Refresh (TRR) mitigations fire once every 72 refresh intervals, delivers effectively 6.6 times the hammering intensity of prior attacks

Four researchers at the University of Toronto have disclosed GPUThor, a Rowhammer technique that breaks the error correction Nvidia has spent the past year recommending as the defense against exactly this class of attack.

Chris S. Lin, Joyce Qu, Aditya Rajeev and Gururaj Saileshwar are expected to present the paper outlining their approach and subsequent findings at ACM CCS 2026.

The attacks target Ampere-generation workstation cards with GDDR6 memory, specifically the RTX A4000, A4500, A5000 and A6000, and it turns an unprivileged CUDA program into a root shell on the host.

An attack that matters much more than its predecessors

Rowhammer works by repeatedly activating a DRAM row until charge leaks from cells in the physically adjacent rows and flips their bits. The attacker never touches the victim's data directly, making it an excellent precursor to tampering, sandbox escape, and privilege escalation, among other things.

This approach is possible despite an in-chip countermeasure called TRR that accompanies ECC-enabled chips on these GPUs. TRR aims to prevent Rowhammer attacks by tracking how frequently specific memory rows are activated and automatically refreshing adjacent rows before a malicious bit flip can occur.

This approach works well on paper and, when Rowhammer attacks hammer uniformly, offers decent protection. The problem arises when approaches such as GPUThor use non-uniform hammering, which is much more likely to succeed. Non-uniform hammering is not exactly new, having already succeeded on the CPU side of the spectrum thanks to the well-documented Blacksmith attack vector.

GPUThor is the third attack from broadly the same group in eighteen months, following GPUHammer in 2025 and GPUBreach earlier this year, and it matters because the previous two stopped working the moment a user typed the command to enable ECC.

ECC was essentially Nvidia's go-to response to reported Rowhammer attacks, but it may no longer be a solution in its current state. By hammering non-uniformly, the researchers opened another attack vector for Nvidia's Ampere-based GPUs, which are overwhelmingly affected by the technique.

With ECC disabled, GPUThor produced 72,000 to 377,000 bit flips per gigabyte across four Ampere-based cards, with Nvidia's A5000 being reported as the most vulnerable. That approaches the roughly 550,000 flips per gigabyte that Blacksmith achieves on DDR4, which is cause for concern: GPU Rowhammer is now in the same league as CPU Rowhammer.

When enabled, the researchers said ECC protectors reduced, but did not completely mitigate, the issue: they delivered double-bit errors and 2 triple-bit errors that the technique 'fixed' by choosing the wrong value.

On an RTX A6000 with ECC enabled, the GPUThor technique forces one GPU reset every 2 hours, killing all running processes on the GPU; as a result, the GPU flags itself as RMA-read within a day.

The team disclosed the attack pattern to Nvidia on 29 April 2026 and subsequently to Google, Microsoft and AWS, and held the work until the 25th of August. A code release is scheduled for the 15th of November, even though there is currently no CVE information or patch being deployed to address the issue.

• Wordfence discloses two critical flaws in Elementor Pro and Super Forms
• Bugs allow unauthenticated arbitrary file uploads, enabling remote code execution; both patched recently
• Exploitation attempts already exceed 440,000

More than six million WordPress users are at risk of website takeover, researchers have claimed after discovering two major vulnerabilities being exploited in the wild.

Security researchers Wordfence disclosed finding two flaws, one in Elementor Pro, and one in Super Forms - two popular WordPress plugins.

Elementor Pro is a commercial plugin that allows users to build websites using drag-and-drop elements instead of code. With it, they can add advanced widgets, templates, different forms, popups, and more. It is quite a popular solution with more than six million websites actively using it.

Two bugs, hundreds of thousands of attacks

According to Wordfence, up until recently, it was vulnerable to an “unrestricted file type upload” bug in all versions up to, and including, 4.2.1. “This makes it possible for unauthenticated attackers to upload files that may be executable, which makes remote code execution possible,” the researchers explained. “This requires that the targeted site has published a page containing an Elementor Pro Form widget with at least one non-required File Upload field.”

The bug is tracked as CVE-2026-32475, carries a severity score of 9.8/10 (critical), and was patched in mid-August 2026. So far, Wordfence alone blocked more than 190,000 exploit attempts.

At roughly the same time, the researchers also reported finding a flaw in Super Forms, a form builder plugin that lets users create and manage forms using a drag-and-drop interface. This plugin, with some 13,000 active installations, contained a bug that allowed arbitrary file upload in all versions up to, and including, 6.3.313.

“This makes it possible for unauthenticated attackers to upload files that may be executable, which makes remote code execution possible,” the researchers explained.

This one is tracked as CVE-2026-14894, also carries a severity score of 9.8/10 (critical) and it, too, was patched a few weeks ago. For this flaw, Wordfence observed more than 250,000 exploitation attempts meaning that cumulatively, these two resulted in 440,000 attacks already.

Given the widespread adoption of these plugins, and the fact that the flaws are being actively leveraged, users are advised to apply the fixes without delay.

Via The Hacker News

Can AI-driven NAS tech be sentimental? UGreen CEO Samuel Zhang seems to think so.

Launching the company’s new smart-home technology suite, he started by recounting a personal story. It was a hot summer’s day, and his family gathered in the backyard for a barbecue. They cooked together, chatted, played, and just had a genuinely fantastic time. But weeks later, when Zhang went to check his photo gallery and relive that cherished day, he was left feeling disappointed. The family hadn't taken any photos.

This, he said, got him thinking about home cameras — devices that are typically used for security. But, Zhang said, “it [a home camera] should see more than just danger — it should also be able to witness the joys of everyday family life”. And this, he said, was one of the key points of inspiration for UGreen’s new suite of local smart-home AI technology.

I got the chance to hear about, and see, UGreen’s home hubs, cameras, digital photo frames and more at IFA 2026 in Berlin. Here’s what I learned — and how the technology could transform the smart home.

Keeping it local

At the heart of UGreen’s smart-home suite is the new HomeAgent. This is an ‘agentic AI local hub' that brings local storage, computing, and control all together in a single device. Through this hub, users can connect up a range of UGreen AIoT devices, such as cameras, smart speakers, and digital photo frames, plus some select third-party products.

A crucial part of the HomeAgent’s appeal is its reliance on local storage and computing. This means that any data that runs through your hub — be that photos, videos, voice inputs, or other files — is processed locally, rather than through a corporate cloud server.

At a time when trust in large corporations is strikingly low, and major data breaches seem all too common, this could be a massive selling point for security-savvy customers. Keeping data local also removes the need for costly Cloud subscriptions, making the initial investment feel that bit more worthwhile.

There are three HomeAgent models scheduled for release. The first is the UGreen HomeAgent HA100, which is the most basic option. It provides all of the core functions you’d need from the hub, but offers slightly less processing power, memory, and storage, among other things, than its siblings. The HA100 Pro is the step-up model, which offers heightened performance, with 205 TOPS (as opposed to the standard HA100’s 26 TOPS), a 10GbE LAN port (compared to the HA100’s 2.5GbE), and a more advanced CPU.

The final boss, if you will, of the HomeAgent trio is the MA100. This is powered by the NVIDIA Jetson Thor T5000, and offers far superior AI performance to the other two models, and considerably more memory and storage.

However, this is certainly reflected in the price, with the MA100 commanding a list price of $19,999, compared to the HA100 Pro, which costs $5,999 and the HA100 at $1,799. Early bird offers are available, which make the HomeAgents about half-price across the board. Shipping is set to commence in early 2027.

AIoT in action

So what does UGreen’s smart home AIoT tech look like in person? Well, pretty impressive.

I was able to check out a demo, and take a closer look at all of the HomeAgent products, and other devices you can connect up to it. UGreen had a HomeAgent connected to a smart speaker, and a camera, among other things.

I tried walking towards a door, and a UGreen smart security camera started blaring after I walked past a certain point, telling me that I was on private property. This data is processed locally through the HomeAgent, so if you ask your connected smart speaker about any unusual activity, it can tell you if anyone approached your door, whether that was a delivery worker, or something more nefarious.

The smart speaker UGreen showed off has a new AI assistant, named Uliya. You can use the smart speaker to do just about anything around your home. You can ask Uliya to remind you of where you were at a specific time, to close the curtains, to play music, or even to identify the location of your pet and cast a live video feed. The functionality you get will vary depending on how invested you are in UGreen’s AIoT ecosystem, but the capabilities are pretty broad.

You can also ask Uliya to give you a summary of captured events throughout the day, display memories on a UGreen smart photo frame in real time, or to show you highlights from a special day — circling back to Zhang’s desire to revive memories that had been lost in time. One function I really liked was the ability to create an Away mode, so that when you step out of the house, lights will turn off, curtains will close, and a robovac will get to work.

UGreen’s new HomeAgents appear to be incredibly capable, feature-rich, and innovative products, and their local handling of data could be game changing for some households. At the product launch, Zhang said “In the future, we hope every family can have a HomeAgent”; and it'll be interesting to see how families adopt such tech moving forward.

• G7 urges governments and organizations to begin transitioning to quantum‑resistant encryption
• Warning highlights risk from future quantum computers able to break current standards like AES
• Guidance: inventory cryptographic assets, prioritize critical systems, and adopt phased PQC strategies

The G7, a collection of some of the world’s most powerful economies, is urging organizations to adopt quantum-resistant encryption as soon as possible and minimize the risk of losing sensitive data to technologically advanced threat actors.

Virtually every industry in the world today relies on encryption (the process of converting readable data into scrambled, unreadable data so that only someone with the correct key can decode and access it) to protect sensitive information. Banks, telecommunications providers, defense organizations, and practically everyone else are using some form of encryption - most probably AES (Advanced Encryption Standard).

Decrypting the protected data without the encryption key is considered almost impossible, since the computational effort necessary to pull it off would simply be too large. However, since quantum computers work on a fundamentally different principle, it is believed that once they are mature enough, they will be able to “crack” today’s encryption standards.

G7's advice

Now, the G7 has urged government and organizations to remain ahead of the curve by deploying “quantum-resistant encryption”:

“To protect themselves from the threat brought by CRQCs (cryptographically relevant quantum computers), organizations should begin planning their PQC transition now and should aim to complete their transitions within any the timelines set out by their national cybersecurity authorities,” the warning reads.

“In addition to quantum-related risks, organizations that delay their PQC transition may lose competitive advantage or may be excluded from contracting opportunities, including public procurement.”

The G7 advises governments and organizations to take a number of steps, including identifying critical systems and prioritizing them. They should adopt a phased and risk-based strategy, start their transition early, and inventory their cryptographic assets. They should also map their dependencies and develop a transition plan.

“To limit transition-related costs, they should opt to purchase products that integrate PQC and replace their systems with quantum-safe ones as part of their standard renewal schedule. As such, starting the transition early could result in lower migration costs overall. Planning and conducting their PQC transition properly would enable organizations to prevent insecure implementations and avoid increased exposure to conventional cyber threats.”

• US military branches disable ad tracking on government‑issued devices after warning
• Location data from personal devices still poses risks; past data remains for sale
• Senators urge DoD inspector general to review policies and stop servicemember data exposure

More than 20 years ago, Jason Bourne was smashing mobile phones to prevent being tracked by his adversaries - and the US Department of Defense (DoD) has realized it should probably do something similar.

In May 2026, Senator Ron Wyden mailed the DoD (among others), informing the agency that foreign powers were using commercial location data to target US troops - and now, Wyden has shared the responses he received from the US Army, Air Force, Navy, Marine Corps, and Special Operations Command, all of which have disabled advertising tracking across their government-issued devices, including Apple and Android smartphones, and Windows computers.

Some branches did it earlier this year, while the Air Force allegedly concluded the efforts in July 2026 - and while Wyden praised the move, he also suggested it is not enough, and that further steps must be taken in order to protect US troops stationed abroad.

Evaluating the leaked files

“We commend these service branches for implementing this cybersecurity defensive best practice on government devices,” Wyden and Senator Pat Harrigan wrote. “However, recent reports regarding the continued availability of commercial location data originating from DoD facilities raise troubling questions.”

Apparently, despite disabling these features and placing the troops out of reach of data brokers and similar commercial entities, the data that was gathered earlier remains available for sale, which still presents a risk.

“The members suggested data could continue to be available for several reasons, including that commercial data available originates from personal devices, not government phones, that are carried by servicemembers and government contractors,” a report on Wyden’s website reads.

“Wyden and Harrigan called on the DOD inspector general to examine commercial location data that DOD and other government agencies have already purchased to determine why existing policies have not prevented servicemembers location data from being sold online, and to recommend policy changes to better protect US personnel.”

Via TechCrunch

• OpenAI hid an incident where a model hijacked a wiki page to use as an AI agent communication board
• The incident was hidden while the company dealt with the fallout of the Hugging Face attack
• The company is now working on a framework for disclosing incidents of 'misalignment'

OpenAI recently disclosed the details of how one of its AI models escaped a sandboxed environment and attacked Hugging Face during an evaluation - and as part of the incident, the models created a messaging board to communicate with each other and influence each other’s reasoning.

OpenAI has now disclosed that shortly after this incident, agents undergoing testing again escaped their ‘secured’ environment and hijacked an obscure German wiki to use as a messaging board. Per Reuters, OpenAI leadership kept the incident hidden while they dealt with the fallout from the Hugging Face incident.

Now that OpenAI has acknowledged its role in the incident, the company has said it is “past time” to put together an incident disclosure pipeline when its models escape testing and slip into third-party networks.

Who is at fault when models do what they’re designed to do?

Before the two incidents, OpenAI said it, “treated misalignment largely as a research question, which gets communicated in research publications”. But now that models are behaving in previously unknown ways and having real-world impacts, the company said it would change its approach “to expand for this new phase of model capabilities”.

The company labelled the most recently disclosed incident as “an instance of misalignment similar” to the Hugging Face breach.

I myself am guilty of reporting on AI breaking out of containment as going ‘rogue’, but these models are doing exactly what they are designed to do. OpenAI’s detailed disclosure of the Hugging Face incident showed that the models were pushed to try and solve a benchmark test by cheating, which is exactly what caused the cyberattack to happen.

OpenAI said that both itself and “the larger AI community do not yet have a clear standard for how to report misalignment that shows up during training, evaluation, and deployment, including examples that don’t look like traditional security incidents but could provide insight into AI behavior and future risks”.

The company added that it is “working on a framework and will share it in upcoming weeks, and in parallel we’re working with dozens of government regulatory agencies worldwide on these issues”.

“When you combine this 'breakout' with the Hugging face breakout, it's starting to display a pattern,” said Ashley Knowles, Lead Cybersecurity Consultant at Black Hills Information Security. “I struggle here with not getting too doomsday-ish but realistically, this is showing a pattern of concerning behavior.”

“I'm wondering if this race to become 'first' is undercutting security measures that need to be taken to properly secure and guard AI agents as they're in development. My concern grows when you consider that OpenAI is also resisting further investigation. Adding onto that, the release and promise that Astra can evade human monitoring. The pot is brewing…”

Ask a CISO why they bought their newest cybersecurity tool, and they’ll have a clear answer lined up.

It stops a specific technique, closes a particular gap, or satisfies a compliance requirement.

However, it's often less clear how the tool fits in with the rest of the stack.

Does it make the overall system stronger, or simply add another layer of complexity to manage?

Research indicates most security professionals already believe they’re juggling too many tools, with over half saying they don’t properly integrate together.

This is a pattern I call ‘additive by default,’ and it results in stacks that grow without a plan, becoming broader but not necessarily deeper or able to match today’s threats.

Start with the outcome, not the technology

I find this additive approach is often due to focusing on specific emerging threats or identified weak points, so decisions are made with a tactical eye rather than a broader strategic view.

Part of the problem is that most organizations have never precisely defined the security outcome they are trying to achieve. Without a shared, specific language for the problem, every new purchase becomes additive, because there is nothing solid to measure it against.

Take asset labelling, which is a core capability most organizations know they need, so they invest in a tool, populate a configuration management database (CMDB), and assign criticality scores. Job done, right? But labelling an asset only answers the question of what it is, and says nothing about how that asset connects to everything around it, or what happens when a policy needs to be enforced against it.

Labelling, visibility and enforcement are three distinct jobs, not one. Solve the first and the second and third remain wide open, so another tool gets bought to cover visibility, then another for enforcement. Each purchase solves its own narrow question perfectly well.

However, none of them were ever asked to work as a single, continuous capability, because nobody defined that as the actual requirement in the first place.

What if asset management labels fed into visibility views and the same reflected how policy is drafted and then enforced? Now you have a strategic problem solved with interoperable capabilities.

How much of your stack is really putting in the work?

Gartner’s most recent Leadership Perspective Survey saw CISOs noting this as a common issue, with just 20-30% of tool capability actually being used in some cases. The instinctive response to this is to cut the stack down, however, that instinct solves the wrong problem.

A good exercise for working out the value of the stack is to evaluate every tool, new or already deployed, against three plain questions.

Does it offer continuous validation against a given threat - and what, specifically, does it validate? Is it still operationally relevant? And is it effective, right now, in the environment you have today? A tool can pass one or two of these and still be failing you.

Virtual Local Area Networks (VLANs) are a good example of this. Twenty or thirty years ago, when networks were typically static and everything likely sat inside a single data center, VLAN-based segmentation was genuinely effective. It matched the environment it was built for.

That environment has since changed almost beyond recognition. Workloads move between on-premises systems, cloud and containers, and nothing stays fixed for long. VLANs are still deployed across many stacks today, still technically doing the segmentation job they were built for. Yet their effectiveness has dropped sharply because they offer none of the continuous validation a hybrid, constantly shifting estate actually requires.

Effectiveness has an expiry date that has nothing to do with whether the tool still runs. Success should be measured by whether the system as a whole still holds up, not by how many tools remain switched on.

Consolidated security, not consolidated tooling

Gartner identified that most organizations are pursuing a vendor consolidation strategy. While this approach certainly reduces unnecessary spending and keeps budgets under control, it’s not necessarily solving the biggest problem.

Consolidated tooling and consolidated security are not the same thing, and it’s an assumption that leads to disappointment. Reducing the number of tools alone achieves little if the underlying processes remain fragmented or teams continue to operate against different objectives.

No data center I have walked through was built entirely by one manufacturer. Racks, switches, storage and cabling come from a dozen suppliers, yet they operate as one coherent system because they were designed to fit together. Security should work the same way.

The goal is not necessarily fewer vendors, but every control, whoever built it, feeding into the same continuous picture of identification, visibility and enforcement.

What to ask instead

Before completing the next security purchase, consider how well a new tool fits into the stack you already have, not just what it claims to do on its own.

A tool that deploys cleanly, validates continuously rather than only at go-live, and feeds its findings back into the tools already in place is doing real work. One that arrives as a fresh, isolated source of alerts is just adding to the noise, however good its individual detection rate looks in a demo. That means comparing what the tool was bought to solve against what it is actually delivering today, checking it against newer capabilities, and being willing to redeploy or renegotiate rather than automatically renew or even retire.

Ultimately, the strongest security programs are not defined by the number of tools they deploy, nor by the number they eliminate. They're defined by how effectively every control works together when it matters most.

We've rounded up the best endpoint protection software suites.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Internet security teams are not short of information. Threat intelligence feeds run around the clock, vulnerability scanners flag thousands of issues a month, and high-profile CVEs dominate the news cycle before most teams have finished their morning coffee.

What is far harder to come by is proof. Raw data is one thing, but what about evidence that the controls sitting in your environment actually detect and respond to the way real attackers behave, in your specific network, today?

That gap is what purple teaming exists to close, and it is where the surprises tend to show up. Running these exercises with organizations that have invested heavily in their security stack, I have repeatedly seen gaps nobody expected, missing telemetry, misfiring detections, and attack paths nobody was watching.

The problem was never simply being under attack. It is being under-validated, and mistaking spend for assurance.

Why so many programs that look mature on paper still fail

Here's a pattern I keep running into with otherwise well-resourced organizations: every control on their books checks out, reports are signed off, tooling is deployed, leadership reassured. None of that holds up once a real attack runs through the environment. The cracks show almost immediately.

Ask where the actual gaps sit, and a familiar list comes back. Some systems feed logs in real time, others barely at all, leaving the SOC with a patchier picture than the dashboard implies. Detection rules are tuned to a generic attacker, not the one likely to show up here. And when something does trigger, it can sit unactioned for hours, since no one owns it.

None of this is exotic, just the ordinary, unglamorous consequence of tools and processes configured once and assumed still to work. With luck, those gaps surface in a simulated exercise before a real attacker finds them.

Part of the problem is what gets measured in the first place. Running standard techniques against a host already flagged as compromised only confirms which rules switch on. That's coverage, nothing more. It won't tell you whether the excess permissions, trust relationships, or misconfigurations in your real environment are exploitable, since that activity looks like normal use and was never built to trigger an alert.

Real validation looks different. It starts with the organization's own risk profile, not a generic library of techniques, and treats configuration and permissions as the real cause, not something a new detection rule can patch over. Only testing your own environment tells you which gap you're actually facing.

What's worth measuring, and the case for continuous validation

Treating validation as an ongoing discipline rather than a one-off exercise also changes the output you’re looking for.

You don’t want another report that sits in a folder until next year's audit; you want to create a living, prioritized backlog. This is a running list of gaps that have been proven to matter, ranked against the paths an attacker could realistically use, not against theoretical severity scores.

It’s an important distinction because a CVE rated critical in isolation may be unreachable in your environment, while a modest misconfiguration sitting on a well-trodden attack path can be far more urgent.

The most valuable information a security team can generate is not a longer list of vulnerabilities, but evidence of which exposures are actually exploitable, mapped against real behavior, and prioritized against genuine business impact.

This is why it is critical to avoid relying purely on agent-based validation. Because an agent is already embedded within the network or on an endpoint, it inherently shortcuts a lot of attack paths and bypasses defensive controls. It assumes the attacker has already achieved a foothold at that specific location, giving you an artificial view of your actual perimeter and lateral resilience.

Alongside the process itself, cadence is another important factor here. An annual or ad hoc test captures a single snapshot, and environments do not stand still between engagements: new services get deployed, configurations drift, staff change. Shifting from ad hoc testing towards an operationalized, continuous cycle of testing, validating, remediating, and retesting is what keeps that backlog tied to the environment as it is right now, not as it was assessed to be six months ago.

The ability to close a validated gap before it can be exploited as an attack path enables the organization to shift from reactive footing to a more proactive stance.

Breaking down the walls between red, blue, and the business

None of this works if validation happens in a silo. The most consistent driver of a strong outcome I have seen is not a tool; it is collaboration between people who are technically on the same side but often operate as if they are not.

Red and blue teams that only meet at the end of an engagement, via a report, tend to treat findings as a scorecard, something to defend or dispute rather than act on. An open-book approach changes that dynamic entirely. When offensive and defensive teams compare notes on what they are seeing, what they expected to see, and why, as the exercise unfolds, gaps get identified and understood together.

If a defensive team cannot follow a technique or lacks the telemetry to see it, that is not a failing to hide; it is the exact information the exercise exists to surface.

Clear ownership matters just as much as clear findings. An alert with no defined owner is functionally the same as no alert at all, so part of building this collaboration is agreeing, in advance, on who acts on what.

That communication cannot stop at the SOC door, either. Validation findings need to travel outward too, translated into terms a business audience can act on: which risks are real, which are theoretical, and where investment will actually reduce exposure. Security teams that build this bridge consistently get faster remediation and fewer arguments about priority, because everyone is working from the same evidence rather than competing assumptions.

Moving towards more coordinated action is another step towards preempting risk, delivering smoother and more effective operations than an isolated response where fixes happen team by team.

Resilience is proven, not assumed

None of this is an argument against investment in tools or people. Rather, it’s an argument for asking a different question that identifies risk. It's time to stop asking, "Are we safe?", because no environment stays safe indefinitely. The question that matters is whether you're resilient enough to validate an exposure and respond before it turns into a breach.

Confidence built on assumptions collapses the moment a real adversary tests it. Confidence built on shared, ongoing validation does not, because it has already been tested, refined, and proven against the way attackers actually behave in your environment.

We've listed the best Linux distros for forensic and penetration testing.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Rhysida ransomware group breached Berlin’s state government, stealing 1.44 million files weighing in at around 5.8TB
• Attackers demanded 30 BTC (~$2.3M); Berlin refused, leading to full leak online
• Leaked data reportedly includes water supply info, staff records, and emergency plans

A cybercriminal group known as Rhysida allegedly broke into the network of Berlin's state government and exfiltrated 1.44 million files. They then tried to extort the government entity for money and when that failed, they leaked it all into the dark web.

According to multiple sources, the group first claimed responsibility for the attack on an underground forum, where it shared a small sample of the stolen files and demanded Berlin pay 30 bitcoin (around $2.3 million) in exchange for deleting the files.

Berlin recognized the attack in a press release, and said it would not be negotiating with the attackers, and instead launched a full-scale investigation into the incident, which it described as an “extremely serious crime and an attack on the state of Berlin.”

Evaluating the leaked files

A few days later, Rhysida decided to leak it all online - and Berlin is now reviewing the leaked files and is assessing the damage:

“After the publication of the stolen data from the Berlin administration, these files are being evaluated at full speed,” a press release, published on the Berlin.de website (machine-translated), reads.

“An additional steering unit has been set up in the Senate Chancellery under the leadership of the Chief Digital Officer (CDO), Florian Hauer, which coordinates the review, examination and evaluation of the leaked data and supports the two Senate administrations concerned in informing and advising the affected citizens and companies.”

According to German public broadcaster Tagesschau, the archive counts 1.44 million files and totals 5.8 terabytes. At the same time, the Chaos Computer Club (Germany's largest and best-known hacker organization) claims the archive contains sensitive data on the city’s water supply, personal data of admin staff, various employment references, and emergency plans.

Via Reuters

• Microsoft reports phishing campaign using ASCII smuggling to bypass spam filters
• Attackers insert invisible characters into keywords, tricking filters and AI agents
• Defenders should normalize Unicode tags and flag unexpected invisible code points as suspicious

Cybercriminals are using the “ASCII smuggling” technique to make sure phishing emails pass security filters and land in people’s inboxes, experts have warned.

ASCII is a character encoding standard that turns characters and words humans can read into numeric values that computers can understand. It can also be used to create characters that aren’t even displayed on the screen (essentially “invisible” ones) but can still be read by the machine.

In a new report, security researchers from Microsoft found crooks are abusing this fact to distribute phishing emails. Most email providers offer solutions that filter out spam emails. These filters look for certain keywords and phrases, such as “funding”, “credit”, “loan”, and similar, and automatically send such emails to the spam folder.

Ongoing campaign

By adding a set of invisible characters in the middle of these keywords, the attackers can break them apart and thus “confuse” the filters.

While the human sees the word “funding” in their email, the security solution is seeing something like “fun[a long string of characters]ding”. This technique has been adopted from prompt injection attacks, where crooks would use ASCII smuggling to deliver malicious and invisible prompts in the emails.

Therefore, when a victim asks their AI agent to summarize the email, it ends up working on a prompt that could be anything from extracting sensitive data, to deploying malware.

Microsoft is saying the campaign has been ongoing for months, and while it peaked in February 2026 with more than 2.3 million emails every day, and has been in decline ever since, it remains active to this day. In early February, Microsoft observed a cluster of almost 150 sender domains, all themed around finance. These domains accounted for almost all (96%) of all the spam emails Defender for Office 365 flagged under ASCII smuggling.

Defending sounds relatively simple, though. IT teams should normalize Unicode tag characters and other invisible code points before applying any keyword detection. They should also consider all unexpected tag-block characters as suspicious.

Via BleepingComputer

You probably have tens, if not hundreds of passwords floating around across all the online accounts you use every week. In fact, the average person has over 150 passwords. It's not easy to remember them all, making it more attractive to keep them simple and similar - but that can lead to a serious chain of account theft.

That's why password managers like Keeper exist. They generate strong passwords and store them passwords in a secured vault to keep them safe from hackers and prying eyes. They even autofill your credentials to make logging in to your account faster and easier.

Keeper has cut prices on its personal, family, and business starter plans by up to 50%. Keeper regularly scores highly in our reviews, and includes great features across Personal, Family, and Business plans.

Why we recommend Keeper

Our Keeper review highlights the platform's zero knowledge architecture and device level encryption that helps keep your password vault secured against unwanted intrusions.

Where Keeper really shone was in our usability and convenience testing. We loved how Keeper uses biometric security to access your vault, rather than requiring a master password with each log in. By using a facial scan or fingerprint to verify that its actually you accessing your vault, it adds a fast but secure method of verification to ensure its you, and not a nefarious actor.

For households, the Family Plan includes five private vaults, allowing you to quickly share Wi-Fi or streaming passwords using Keeper's shared vaults.

The Business Starter plan covers 5 users, covering an unlimited devices, shared team folders, user activity reporting for IT admins, and a free Family Plan for every team member.

These discounts apply to the first year only, so the value is highest for new customers or anyone switching from a monthly plan. After that, pricing goes back to standard rates.

• Chinese companies have reportedly hidden their ownership to operate in Taiwan
• They're also poaching Taiwanese workers, paying them up to 10x more
• China is fighting back against US restrictions, but Taiwan wants to protect its own industry

Taiwan has reportedly increased the pressure on Chinese technology companies that are hiding their Chinese ownership as it tried to protect its own semiconductor business.

The country's Ministry of Justice Investigation Bureau (MJIB) has reportedly investigated 166 cases involving suspected illegal Chinese ownership since 2020, resulting in 36 convictions – 33 of which involved in semiconductor research, development or design.

While Taiwanese law doesn't prohibit Chinese companies from operating in Taiwan, they must seek government approval under the Cross-Strait Act, but by obscuring their Chinese ownership they've been looking to circumvent those restrictions.

Chinese companies targeted for operating illegally in Taiwan

Rest of World reports many of the Chinese firms in question established shell companies in third countries or used non-Chinese nationals to register their businesses, making them appear either Taiwanese or generally non-Chinese.

The news comes against a backdrop of heightened geopolitics – China has been investing heavily in growing its semiconductor self-sufficiency amid ongoing US restrictions, and have reportedly been poaching Taiwanese engineers with salaries up to 10x higher than average.

Taiwan itself is said to account for more than 60% of global semiconductor production, producing around nine in 10 of the world's most advanced processors, so alleged Chinese poaching is especially noteworthy. "Taiwan’s high-tech talent has become a key target for recruitment by Chinese enterprises," the MJIB wrote.

In a separate case, the MJIB has also investigated 67 cases involving Chinese companies suspected of stealing Taiwanese trade secrets.

Interestingly, the cases don't just relate to smaller start-ups – Xiaomi, OnePlus, Oppo and Semiconductor Manufacturing International Corporation (SMIC) were all named in the report.

Fines reported in the cases ranged from around $3,800 to $800,000, with penalties of up to three years' imprisonment also cited.

In its latest financial stability report, the Bank of England warned that an AI crash could plunge the UK into recession, claiming that a “price correction in AI stocks, driven by a change in productivity and profitability among tech-led companies, could cause a 2.2 percent fall in U.K. GDP.”

This should serve as a warning to businesses everywhere.

The risks surrounding AI extend far beyond Silicon Valley.

As leaders prepare for what comes next, however, many are still asking the wrong question.

“Is AI real, or is it hype?”

The last 30 years of technology should have taught us that the answer can be both.

The last tech crisis of the 20th century

Y2K was a real technical risk that came to look, in hindsight, like mass overreaction.

The UK spent billions preparing for the millennium bug, with organizations testing and patching systems that underpinned everything from banking and benefits payments to air travel and the National Grid.

The government established Action 2000 to help businesses prepare, while banks and major infrastructure providers ran extensive contingency plans ahead of the deadline.

Then midnight passed, the planes stayed in the sky, cash machines kept dispensing money and the lights stayed on. The whole affair began to look almost laughable.

Yet the apparent anticlimax obscured an important point. Much of the disruption people feared was likely avoided because organizations took the risk seriously and prepared for it. Cybersecurity has the same problem.

A threat that is successfully mitigated can look, in hindsight, remarkably like a threat that was exaggerated.

Not to burst your bubble

The dotcom bubble taught a different lesson. The internet was not fake, but the valuations often were. Boo.com, for example, raised vast sums to reinvent fashion retail online, only to collapse in 2000 after burning through its funding. Companies with little profit and, in some cases, barely coherent business models were treated as if they had risen above the laws of economics.

Then the bubble burst, fortunes disappeared and many of the supposed pioneers vanished, but the ideas they were betting on did not. In fact, I bet you’re reading this on the internet right now.

Style over substance

Crypto added a third lesson. Revolutionary language can hide weak use cases. There were serious technical ideas beneath it, but the public boom became dominated by speculation, celebrity endorsements and the strange habit of starting with an asset and then searching for a purpose.

At the height of the NFT frenzy, an NFT of Twitter founder Jack Dorsey’s first tweet sold for $2.9 million. A year later, its owner attempted to resell it for nearly $48 million, only to see initial bids fall spectacularly short. The technology had not disappeared. What had disappeared was the market’s willingness to treat digital scarcity as lasting value.

AI’s rise carries elements of all three stories.

A familiar visage

Like Y2K, AI presents risks that may sound exaggerated until they are not. As any cybersecurity expert will tell you, if the worst harms are prevented, the warnings may later look hysterical. This will not prove they were imaginary.

For AI, this means treating cybersecurity and governance as operational disciplines, not innovation-team side projects. Organizations need to understand what data is entering AI systems, what those systems can access, what authority they have, how they could be manipulated and how the business would respond if they failed.

With the dotcom boom, AI looks like a real general-purpose technology wrapped in an overheated investment story. It can already write code, summarize documents, and automate parts of professional work. But that does not mean every AI company is valuable, every AI product is useful, or every data center will earn its keep.

For businesses learning from this, they must invest in capabilities, data foundations, people and workflows, rather than betting everything on whichever vendor, model, or product is making the most noise this quarter.

And like crypto, AI has acquired a mythology. “AI-powered” is becoming what “blockchain-enabled” briefly was, a phrase that can mean something, nothing, or merely “please value us higher.”

For the adoption of AI, business leaders must now ask boring, but necessary, questions. What problem does this solve? Who uses it? What data does it see? What systems can it influence? What authority have we given it? What happens if it is wrong, manipulated or unavailable? And who is accountable when that happens?

This combination is what makes the current moment so hard to read. The lazy argument says AI is either a revolution or a bubble. History suggests a more uncomfortable possibility; it could be both.

“The Spirits of all three shall strive within me. I will not shut out the lessons that they teach.”

Too often, business leaders are trying to win an argument about whether AI is overhyped when they should be building a strategy that survives either answer. The sensible response to AI is neither a moonshot nor a moratorium. It is disciplined experimentation.

This means treating AI less like a campaign slogan and more like a portfolio of bets. Some of those bets should be defensive, focused on understanding where AI is already being used, what data and systems it can access, and whether the organization can respond when things go wrong.

Others should be exploratory, testing where AI can create genuine value across areas like coding, customer support and knowledge management. And some should be deliberately skeptical, particularly when tools sound impressive but cannot demonstrate savings, better outcomes, or clear ownership.

If the aim is simply to have an AI strategy that reads well in a board pack, then the lessons of the past 30 years have been missed. Y2K, dotcom and crypto each showed us that hype and substance are not opposites. Real risks can be exaggerated. Transformative technologies can attract irrational investment. Powerful ideas can coexist with weak use cases.

AI may prove to be all three at once. The businesses that navigate it best will not be those that predicted the future perfectly. They will be those that were prepared to adapt when reality arrived.

Best cloud storage: tested, reviewed and rated by experts.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Around three in four UK medical practices use Docsign to share important letters
• Docmail has been down for five days due to a server issue
• Services can be resumed once the investigation completes

Docmail, a cloud-based mail service used by high-profile businesses and public-sector organizations, including some NHS trusts, continues to suffer a major outage after first being hit on September 2 2026.

The platform works by customers uploading their communications digitally, with Docmail handling printing and postage to recipients.

Parent company CFH says more than 75% of UK medical practices use its services for bulk and individual patient mail, spanning GPs, dentists, pharmacies, opticians and other private-sector firms.

Docmail is still down

Now on its fifth day of interruptions, it means healthcare appointment letters, council correspondence and other important mail could be missed or delayed.

Docmail said customers were experiencing "slower than usual processing times" on September 2 before taking the service offline for "essential maintenance."

On September 4, the company told The Register: "We have been made aware of an issue impacting an isolated server, used by only one client." In other words, CFH purposely shut down the wider platform to investigate this incident, however the finer details of what went wrong are under wraps.

"This process does take time, and we are acting on the advice of external consultants to ensure we are taking the right precautions," a company spokesperson added.

According to CFH, Docmail itself is ready to be restored and made available, but external reviewers are still conducting the investigation.

The company's status page has not been updated since September 4 at the time of writing – at that time, it labelled the outage as an "unresolved incident."

"Work remains ongoing as a priority, and we will provide a further update as soon as we have more information to share," the company added.

• Nvidia pays $11.9 billion for Hugging Face, and a further $1 billion in equity-based retention
• The company promises to keep Hugging Face "an open platform for the entire AI ecosystem"
• Nvidia CEO Jensen Huang has publicly supported open weight models

Nvidia has confirmed plans to acquire Hugging Face for $12.9 billion subject to regulatory approval, marking a major expansion for the chipmaker as it snaps up one of the most important AI model distribution platforms.

In a post, the company asserted it plans to keep Hugging Face as it is: "Hugging Face will remain an open platform for the entire AI ecosystem," CEO Jensen Huang wrote.

Huang explained Nvidia's "infrastructure, engineering and global reach" will help improve Hugging Face's "reliability, safety, model evaluation, inference and deployment capabilities."

Nvidia confirms Hugging Face acquisition

The platform currently counts more than 18 million developers, researchers and creators as users, with more than three million models shared on it.

"Hugging Face will continue to support open source and open weight models from across the ecosystem, from every model builder," Huang stressed again, noting his public backing for open weight models. Ordinarily in an acquisition, questions would be raised over whether, in this instance, Nvidia's own GPUs and software could receive preferential treatment, but the company is doing everything it can to assure us that won't be the case. A pending regulatory investigation will also ascertain this.

Nvidia also claims to be the biggest contributor to Hugging Face, having shared more than 500+ models and around half as many open datasets.

The news comes less than two weeks after reports started to circulate about a potential acquisition, and if it closes, it'll be Nvidia's biggest acquisition to date. The $12.9 billion payout comprises a Hugging Face valuation of $11.9 billion (way above its earlier $4.5 billion valuation in 2023) and an equity-based retention program of around $1 billion.

As for upcoming regulatory interjection, Nvidia warned in a SEC filing that "Government restrictions may negatively impact our business and the Hugging Face platform."

• 2027 UBS recruits told they must have AI proficiency and a willingness to learn
• The bank's AI Fluency Pathway will continue to support junior workers' development
• While thousands of jobs could be at risk, we're starting to see shifts rather than outright displacement

Swiss investment giant UBS is now requiring all junior bankers to demonstrate AI proficiency as the skill moves from being a nice-to-have to an absolute requirement within recruiting.

The change currently applies to graduates and interns applying to the company's 2027 intake, per the Financial Times, and it's unclear whether UBS will broaden the requirement to all workers in the future.

As part of the new requirement, recruits will need to be able to demonstrate that they can use and experiment with AI responsibly to improve business outcomes – not just that they can use popular AI chatbots like ChatGPT.

UBS makes AI skills a must-have

AI-related questions will now become part of the bank's recruitment interviews on top of both the existing types of questions as well as the usual requirements, like a 2:1 degree.

While the news puts additional strain on graduates who now need to invest in their own AI skills, it's an example of how artificial intelligence isn't replacing entry-level workers, with the bank seeing it more as a productivity booster for human staff.

UBS' training program will also include an 'AI Fluency Pathway' to cover real-world banking AI use cases and responsible AI use, implying that the bank is more focused on prospective workers being able to prove a certain level of proficiency and willingness to learn – not full proficiency from the get-go.

AI's longer-term effects on banking employment are more unpredictable, though, with an earlier Morgan Stanley report warning that 200,000 banking jobs could be lost in Europe over the next five years. That was in early 2026.

While the outlook now seems more positive that junior workers may not be at a loss, it's clear that roles are evolving and entry-level workers could see their responsibilities shift toward AI management.

For the last two years, the advice to brands has been more or less standardized: get cited and recommended in AI, and the rest will follow.

Whether that is called GEO or AEO, the objective is the same: make sure AI can understand your brand, retrieve the right information and recommend it when a customer asks.

That was until Amazon quietly published a number in its Q2 results that starts to undercut this advice.

Andy Jassy revealed that shoppers who click a paid Sponsored Prompt inside Alexa for Shopping convert to a sale 48% more often, and spend 21% more, than shoppers who do not.

It is Amazon's own data rather than an independent industry benchmark, but it is one of the clearest signals yet that paid placement is becoming native to the AI shopping experience, rather than simply sitting alongside it.

That matters because AI shopping is beginning to split into two very different models: open assistants that aim to surface the best products they can find, and closed ecosystems that control the commercial environment around the recommendation.

For brands, those models create very different ideas of what visibility is worth.

Why are AI shopping platforms pulling apart?

Until recently, it was reasonable to talk about AI visibility as one thing. A brand wanted to be understood by the model, cited by the assistant and recommended when a customer asked a relevant question. That assumption becomes harder to sustain when the platform making the recommendation also has an advertising business to monetize.

Amazon is the clearest example. A Sponsored Prompt can appear inside the same conversational journey in which a customer is deciding what to buy. Google is moving in a similar direction, bringing advertising into increasingly conversational search and shopping experiences.

The commercial incentive is obvious. If a paid recommendation inside an AI conversation converts better than a conventional ad, platforms have a reason to put more advertising into that conversation.

Open assistants face a different calculation. Their value depends heavily on the perception that recommendations are being made because they are relevant, rather than because someone paid for them.

That creates a much harder balance between monetization and trust. The result is not one AI shopping channel, but multiple ecosystems developing around different commercial incentives.

Is GEO/AEO still enough?

The fundamentals behind GEO and AEO are not going anywhere. Accurate, specific and well-structured information still gives AI systems a better chance of understanding a product and deciding when it is relevant.

But there is an important distinction emerging: GEO and AEO solve the visibility problem. They do not necessarily solve the commercial problem.

A brand can be highly visible in an AI recommendation and still fail to convert that visibility into revenue. Equally, paid visibility cannot compensate indefinitely for poor underlying product information. An AI still needs reliable data about what a product is, who it is for and how it compares with alternatives.

The question for brands is therefore becoming bigger than simply whether they are being recommended. They need to understand how recommendation works on each platform, what happens when advertising enters the same decision-making process, and whether their visibility ultimately leads to customer acquisition.

Where does ACO fit?

This is where agentic commerce optimization, or ACO, starts to become relevant.

GEO and AEO are fundamentally about getting a brand understood, retrieved and surfaced by AI. ACO takes that problem into the commerce layer, where an AI is beginning to make or influence the purchasing decision.

For an agent, product content is only part of the equation. Price, availability, specifications, variants, delivery, returns and the ability to complete a transaction all matter. A product can therefore be well optimized for AI visibility and still be a poor choice for an agent if the information it needs to act is incomplete or inconsistent.

That is why ACO can be thought of through the 5Cs: Completeness, Context, Citations, Correctness and Customer Acquisition.

The first four help determine whether an AI system can confidently understand and recommend a product. The fifth asks the commercial question that visibility metrics alone cannot answer: did that recommendation create a customer?

ACO is therefore the operational response to the next stage of AI shopping: continuously monitoring how products appear across AI platforms and fixing the underlying content, catalogue and commerce data when those systems start to drift.

That becomes more important as platforms develop different approaches to advertising and recommendation. A brand cannot optimize for a single version of AI shopping when the underlying platforms are making different commercial bets.

What should brands do now?

The answer is not to abandon GEO or AEO in favor of another acronym. The fundamentals remain the same: accurate product information, clear structure, consistent data and content that answers the questions customers actually ask. The difference is that brands now need to monitor what happens beyond the initial recommendation.

Which products are AI platforms surfacing? Which competitors are gaining ground? Where is product information inaccurate or missing? How does that change between assistants? Where does paid visibility enter the journey? And, ultimately, is that visibility generating incremental customer acquisition?

Amazon's 48% conversion figure matters because it suggests that appearing inside an AI conversation at the point of purchase intent can be commercially different from appearing alongside one. If other platforms follow, the distinction between AI visibility, paid media and commerce operations will become increasingly difficult to maintain.

The opportunity is therefore not to replace GEO or AEO, but to build on them. As AI moves from answering shopping questions to influencing purchasing decisions, visibility becomes the starting point rather than the end goal.

The brands that take advantage of this state of play will be those that combine GEO/AEO fundamentals with ACO - optimizing the five Cs and, ultimately, measuring AI not just by how often it mentions them, but by how much incremental revenue it helps generate.

We list the best data visualization tools.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

As the US moves to integrate more AI systems into every arm of its military, the situations on the ground in the Middle East are raising some serious problems for maintaining the compute and connectivity needed to keep systems online.

Drones, sensors, AI-assisted targeting and decision-making all rely on these systems in one way or another, and the destruction of a single frontline AI data center can seriously damage an army’s capacity to function.

But these problems go beyond hardware. How does a drone adapt its mission parameters without being able to communicate? How can AI models continue inference without access to critical battlefield data? And what compromises must be made to prepare battlefield AI for these conditions?

Battlefield AI should continue functioning, even when comms go down or hardware is disabled

While the conflict circling the Strait of Hormuz may be a hugely expensive endeavour, Iran is providing valuable lessons for warfighters. Relying on traditional data centers in the Middle East has already shown its weaknesses. These sprawling warehouses make easy targets for cheap missiles, and entire systems can be taken offline from a single hit.

Palantir has already begun deploying shipping containers filled with Nvidia hardware as an alternative to the traditional data center. These containers can be deployed on frontlines and can connect directly into operational workflows, offering decentralized compute where it is needed most.

Scaleout is a company building infrastructure for edge AI and federated learning that trains models on distributed, sensitive data without centralising it. The company has worked with NATO - most recently the Swedish Air Force - and defense primes such as BAE Systems.

Andreas Hellander is the CEO and co-founder of Scaleout. He holds a PhD in Scientific Computing and an MSc in Biotechnology Engineering, and is an Associate Professor at Uppsala University, where he built one of the top research groups at the Department of Information Technology before founding Scaleout.

I spoke to Andreas to learn more about how battlefield AI systems can be made more resilient, and how drones, sensors, and other frontline technologies can continue functioning when conditions go downhill.

• What happens when parts of a battlefield AI's computing and hardware are taken offline? What redundancies are in place to keep communications between systems operational?

The aim is not to promise an unbreakable connection; it is to stop the loss of one component from taking down the whole capability. In our architecture, a disconnected node continues local inference using its last approved model, caching detections and telemetry until a link returns.

Multiple aggregation points can be used above the edge. If one becomes unavailable, clients can be reassigned and training can proceed with those still reachable. That creates graceful degradation rather than an all-or-nothing failure. We do not supply the tactical communications network itself; our job is to ensure the AI workload does not assume that any radio, satellite or other link will always be available.

• What are the trade-offs between a highly centralized battlefield AI network and a more distributed approach, especially concerning data gathering, targeting, battlefield coordination and logistics?

Centralisation provides a consistent operational picture, substantial compute and a clear place to enforce security, model approval and command policy. It supports theatre-wide coordination and logistics, but also concentrates risk: losing one link or facility can remove a disproportionate share of the capability, while moving large or classified datasets may be slow, prohibited or impossible.

A distributed system keeps analysis close to the sensor, reducing latency and allowing local functions to continue during disconnection. The trade-off is fragmentation. Targeting based on a partial or stale picture is risky, and local logistics decisions may not be optimal for the wider force. The best design is usually hybrid: central coordination when available, with clearly bounded local capability when it is not. Authority remains with the relevant command-and-control and human decision process.

• How do these AI drones and sensors adapt to the battlefield and what differences are there in the behavior of an AI drone that can still communicate and one that has lost communication?

It is important to separate running a model from retraining it. Small platforms typically perform inference in flight, while controlled retraining and approval happen on a ground node or more capable edge computer. Onboard learning exists in research, but uncontrolled mid-mission retraining creates serious assurance problems.

A connected platform can share detections, receive tasking and contribute to a wider operational picture. After losing its link, a drone follows its pre-approved failsafe: depending on the mission, it may continue a bounded task, return, hold or land. Onboard processing can keep perception and navigation working, but the platform loses external context, new instructions and fleet coordination. Disconnection should be a designed operating state with explicit limits, not an improvised emergency mode.

• How is the data collected by these drones and sensors used to help battlefield AI systems adapt to new conditions, threats, and tactics?

The useful data is often the hardest to move, so we send the model to the data rather than centralising the footage. Active-learning software identifies the frames most likely to improve it, such as uncertain detections, unfamiliar objects or changed conditions. A person still labels the selected material; the software reduces the searching, not the need for human judgement.

The model is then fine-tuned locally and tested against a benchmark. A candidate can also run beside the current model in shadow mode before an operator approves it. Where policy and connectivity allow, sites exchange protected model updates rather than raw training footage, combining them into a shared model. Federated learning reduces data movement, but the updates still need to be secured, validated and audited.

• What are the main challenges you encounter when deploying AI drones and sensors? What capabilities are available for militaries to source the required compute and energy production close to the front lines?

The hardest constraints are usually power, heat, weight, weather, vibration, electronic warfare, inconsistent data and safe software updates across a mixed fleet. Integration and accreditation can be as demanding as the machine learning.

The compute is commercially available: low-power modules such as NVIDIA Jetson for small platforms, and ruggedised GPU servers or containerised edge units for ground sites. Energy is tougher. Every processor competes with propulsion, sensing and communications, while generators, batteries and distribution equipment add logistical burden and new points of failure. In our public demonstration with BAE Systems Bofors, the perception workload ran on onboard compute at -18°C without connectivity. The principle is to use the smallest practical footprint and assume both power and bandwidth are scarce.

• There is an increasingly blurred involvement of commercial companies providing services for military use, such as AWS data centers in the Middle East and drone production in the UK for Ukraine. How are companies providing services to the military adapting to the pressures of potentially being deemed a legitimate military target?

Speaking for Scaleout, rather than for the sector as a whole, we design our platform so that a customer’s operation does not depend on Scaleout remaining continuously available. It runs in the customer’s own infrastructure, works air-gapped, and does not require a licensing call-home or persistent connection to our systems or a public cloud.

That matters in any contested or disconnected environment: reliance on a remote provider can become an operational vulnerability. Our approach has therefore been to keep control and continuity local to the customer, rather than make our own infrastructure a critical dependency.

The wider question of how commercial providers are viewed in a conflict is one for policymakers, legal experts, and the companies involved. We would not want to speculate about other providers’ risk assessments or operating models.

• What is the future trajectory of these systems? Will each individual drone and sensor have its own connectivity that can feed data back through a global satellite network? Will these systems be able to operate autonomously for multiple days or weeks without human input?

Connectivity will improve, including through low-Earth-orbit satellites, but connecting every sensor directly to space is unlikely to be the universal answer. Terminals consume power and spectrum, add cost and can create an electronic signature. A more resilient pattern combines local mesh links, opportunistic synchronisation and selected gateways to satellite or other long-range communications.

Longer disconnected operation is realistic, but endurance is often limited by batteries, weather and maintenance rather than AI software. There is also a critical distinction between operating without a connection and operating without human authority. The first is a resilience property; the second is a policy, legal and ethical decision. The future is therefore not simply more autonomy, but verifiable control: knowing which model ran on each platform, what produced it, where its limits were and who approved it.

OpenAI has unveiled a much anticipated AI model which the firm has dubbed ‘GPT-6 Astra’. While the model has improved significantly across benchmark testing and brings a host of new business features, there is still a dark cloud looming over the new model.

Off the back of OpenAI’s accidental hack of Hugging Face and the company’s subsequent efforts to improve how AI agents behave and interact, numerous cybersecurity experts have raised concerns about the model’s new ‘recurrent depth’ reasoning capabilities.

This new reasoning architecture allows the model to consider a problem multiple times before taking an action, compared to the standard chain-of-thought reasoning used in previous models.

Why the concern about recurrent depth reasoning?

This new level of reasoning apparently offers improved performance. OpenAI also says it has fixed its models' abilities to circumvent boundaries when performing tests by monitoring the models reasoning and ensuring the model stays aligned within the scope of its task.

During Astra’s launch event, OpenAI chief scientist Jakub Pachocki said: “We will not accept degradation in our ability to monitor model alignment beyond a certain level. We will withhold scaling until we can regain enough confidence.”

But numerous experts believe that the lessons of the Hugging Face incident have not yet been learned, and the model has been released without adequate testing on Astra’s reasoning and monitoring.

After all, no one thought one of OpenAI’s models could set up a hidden internet-connected messaging board that allowed AI agents to influence each other's behavior.

But with Astra being released into the real world, the lessons may have to be learned on the fly.

Expert perspectives on OpenAI Astra release

• James Blake, VP of Global Cyber Resiliency Strategy at Cohesity:

The launch of Astra is raising questions again around the safety of Frontier AI. Instead of simply asking whether a model is "safe", organisations now need to ask whether it remains safe across millions of different situations, prompts and interactions. Cyber resilience has traditionally assumed that systems and threat actors behave deterministically. AI systems don’t.

Advanced models can and will continue to exhibit behaviours that emerge from their optimisation process rather than from explicit programming. We have to move beyond thinking about AI as just another software tool and find ways to ensure these systems remain observable, auditable and governable throughout their lifecycle.

The most important question we’ll need to answer in future is one of liability. Suppose an AI system autonomously develops a strategy that causes financial loss, leaks confidential information or violates regulation. Who is responsible? The developer that trained the model? The cloud provider operating the infrastructure? Currently the answer is surprisingly unclear. It’s not just about what AI can do: it’s about who is accountable when it does something nobody expected.

• Oleksandr Yaremchuk, Co-Founder & CTO at Manifold Security:

OpenAI is calling Astra its most aligned model yet, even as its chief scientist admits monitorability is getting harder as models get more capable. Evidently, Astra hides its reasoning in the majority of tested cases, and some successful attacks left no reasoning trace at all. That's the tool many organisations still use, including the labs themselves, for auditing what an agent is doing, and it's getting less reliable with every release.

That matters because Astra isn't staying inside OpenAI's test environment. It's going to run as an agent on employee laptops and in the browser, holding real credentials, inside companies that have no way to watch what it does once it's there. A model that explains itself less isn't more aligned, it's just harder to catch when it goes wrong.

Labs can keep debating what these models say or refuse to say. Security teams need to stop relying on that and start monitoring what agents actually do at runtime, with the ability to shut one down mid-action. That's the only oversight left that still works once the reasoning goes quiet.

• Kristin Lowery, Field CISO at Optiv:

For boards and executive leaders, the emergence of OpenAI’s Astra model highlights a broader reality: AI is no longer just a productivity issue; it is a risk management issue.

Just as organizations established governance for cloud adoption and digital transformation, they now need clear policies, strong oversight, and accountability for AI use.

The question is not whether AI will become more capable — it will. The real challenge is whether organizations can strengthen their governance, security controls, and workforce readiness quickly enough to keep pace.

• Patricia Titus, Field CISO at Abnormal AI:

OpenAI crossing this threshold deserves attention. Credit where it's due, they're handling it responsibly by restricting Astra's advanced cyber capability to a small coalition rather than releasing it broadly.But this isn't one company's problem to contain.

Once a model can find and exploit unknown flaws without a human in the loop, that capability doesn't stay exclusive for long. Open-weight and modified models typically trail the frontier by only months, and that's the reality defenders have to plan around now.

Static, signature-based defences were built for attacks that repeat. They weren't built for an adversary that generates a new one every time. Defenders need the same shift, systems that learn what normal looks like for every identity, human, machine, or AI agent, and flag and contain the moment something deviates, at machine speed.

The window to build that is open now. It won't stay that way once this capability is common instead of rare.

• Raghu Nandakumara, VP of Industry Strategy at Illumio:

With the Astra announcement, OpenAI is doubling down on monitoring the model's own behaviour – a response to the model "breakouts" seen over the past few months.

When Anthropic announced Claude Mythos Preview, the core concern was the model falling into the wrong hands. OpenAI's answer goes further adding guardrails around the model's own reasoning and actions, regardless of the user's intent. The goal is to catch a model going rogue mid-task, not just stop it being misused at the outset.

The rest of this announcement can be summarised as ‘we have a new frontier model, and it’s more capable than the last one’.

• Nvidia's NVHBM moves the memory controller off the accelerator die and into the HBM stack's base die, allowing it to claim a potential 30% more bandwidth, 15% lower HBM power, and 25% more compute area versus standard HBM4E
• Nvidia's NVHBM performance comparisons are linked to HBM4E, which is still in the sampling stage, with a launch expected sometime in 2027
• Access to the custom solution is gated behind NVLink Fusion; Amazon's Annapurna Labs is the only named partner so far and has not said which GPU uses that memory

On August 26, Nvidia extended its NVLink Fusion program with NVHBM, a custom high-bandwidth memory architecture that the company says delivers up to 30% more memory bandwidth per stack, 15% lower HBM power consumption, and up to 25% more usable area on the accelerator die.

The performance was measured against standard HBM4E samples even as the memory is expected to ship in volume some time in 2027.

Amazon's Annapurna Labs is the first named partner for Nvidia's memory advance which aims to address growing performance limitations for frontier-level AI models centering around limited bandwidth.

What makes NVHBM tick?

Nvidia's approach doesn't reinvent how memory is handled; it reimagines where it is managed. Conventional HBM approaches split this between two companies: the memory vendor supplies the DRAM stack and its base die, and the accelerator designer puts the memory controller and physical interface on its own compute die.

JEDEC standardizes the connection between the two at the cost of a very wide, comparatively slow parallel bus. NVHBM dissolves the seam altogether by moving Nvidia's memory controller into the base die of the 3D stack, replacing the standard bus with a narrower, serialized die-to-die link that Nvidia designs and memory makers can build.

Nvidia's technical blog puts the resulting reduction in interface and support area at up to 67% against the JEDEC HBM4E standard, even as it improves key power consumption, memory bandwidth, and usable area.

The underlying tech is not completely new, however: Marvell announced the same basic idea in December 2024, naming Micron, Samsung, and SK hynix as collaborators while claiming up to 25% more compute area, 33% greater memory capacity, and a 70% reduction in memory interface power, figures that come close to what Nvidia is currently projecting. Counterpoint Research analyst Neil Shah put it plainly: "The technology is not new. The distribution is."

Even as HBM4E remains elusive, as it is not in mass production just yet, with Samsung shipping the first HBM4E samples in late May 2026 and SK hynix pulling its own sampling forward to around June, NVHBM is even more elusive. It is a building block gated behind NVLink Fusion, Nvidia's program for connecting third-party accelerators to its rack-scale platform, and access runs through that program's partner list.

Amazon's Annapurna Labs is the first named participant, with Nvidia's blog saying Annapurna will support NVLink Fusion with Trainium4 without explicitly mentioning NVHBM.

For now, NVHBM might be the future, but it may arrive well after HBM4E is already being deployed en masse. Nvidia's own technical blog describes NVHBM as built on the same technology the company will use for future GPUs, without specifying the first generation that would support it natively.

• Google released Chrome 152.0.7977.82/.83 for Windows, Mac, and Linux
• Patch fixes 12 flaws, including zero‑day CVE‑2026‑85046 (type confusion in V8)
• Sixth Chrome zero‑day this year; Chromium browsers also affected, update urged immediately

Google has released a new version of its Chrome browser for Windows, Mac, and Linux, fixing a dozen of vulnerabilities. Among them is a high-severity flaw that is being actively exploited in the wild.

In a security advisory published on the Google blog on September 3, the search engine company said Chrome’s newest version is now 152.0.7977.82/.83 for Windows and Mac, and 152.0.7977.82 for Linux.

Rollout is expected to be gradual, so make sure to check if you already received it. Most of the time the update is automatic and instant and if you’re not certain, tap the three vertical dots in the top-right corner of the browser and choose Help. Navigate to About Google Chrome and there you will find the version number.

Running malicious code remotely

This patch fixes a total of 12 vulnerabilities, most of which are graded as high severity. Among them is a “type confusion in V8” bug, discovered by security researcher Salvatore Gulizia.

For his effort, Gulizia was awarded $1,000. Apparently, this bug is being actively leveraged in real-life attacks, although Google (as usual) decided not to share the juicy details until the majority of browsers are protected.

The vulnerability is tracked as CVE-2026-85046. On the National Vulnerability Database (NVD), it is described as a “type confusion in V8 in Google Chrome [that allows] a remote attacker to execute arbitrary code inside the sandbox via a crafted HTML page.”

V8 is Chrome’s JavaScript engine which allows web apps and interactive websites to run in the browser. Those interested in the technical breakdown of the vulnerability can find it on Guzlia’s blog.

You can find the full list of fixed vulnerabilities on this link.

This is the sixth zero-day Google fixed in Chrome since the start of the year. It also affects other browsers built on Chromium, so if you’re running Edge, Brave, Opera, or Vivaldi, make sure to update to the latest version as soon as possible.

Via The Hacker News

• OpenAI rolled out GPT‑6 Astra, its most advanced AI, with “critical” cyber capabilities
• Astra can autonomously find and exploit unknown flaws; release limited to vetted Daybreak participants
• Staggered rollout sparked user frustration; Altman apologized, promising broader access soon

OpenAI has begun rolling out GPT-6 Astra, its latest and most capable artificial intelligence model.

In an announcement published on its website, OpenAI said Astra represents a “significant step up in cyber capabilities”, meeting the company’s “Critical” threshold.

This threshold, the company explained, means that under the right circumstances, the tool can find “previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step.”

Apologies for a messy rollout

OpenAI also said Astra improves on functions such as computer use, software engineering, and science, and that it should be better at staying oriented, compared to previous versions.

Just like many of the previous models, Astra’s release has been marred with controversy. Roughly a month ago, OpenAI said the model’s “critical” cybersecurity capabilities prompted it to pause some internal development and trigger safety protocols.

In response to early findings, the company scaled up certain security controls and even paused activities that do not meet its newly established security requirements.

It now decided it was time to release the model, but still not to the general public. Instead, it opted for a staggered release, giving it first to a limited group of companies participating in the Daybreak program, a cybersecurity initiative that gives vetted organizations (cybersecurity and otherwise) specialized AI tools for vulnerability research and threat detection.

This did not sit well with many ChatGPT users, especially those paying a monthly subscription. As The Verge reported, CEO Sam Altman was apologizing, mere hours after launch, for the “messy rollout” of Astra.

Initially, all Plus, Pro, Business, and Enterprise users were supposed to gain access to Astra, as well as those accessing through OpenAI API, Microsoft Azure, and AWS Bedrock.

“We are working towards getting Astra in everyone’s hands as quickly as we can,” Altman said on X. “I know it is frustrating and I appreciate the patience. It should be quick.”

Via CNBC

• Texas has gone from being pro-data center to anti-expansion
• Communities have had enough of rising prices and strained supplies
• AI data centers have been described as this election season's core fight

Republican politicians Texas are starting to turn against the rapid expansion of AI data centers, marking a major shift in attitude against backing Trump's push for the US to dominate the global AI race.

Candidates from the party in at least five states have recently shared stances that counter the current administration's goals as citizens continue to be concerned about impacts like rising bills.

Nowhere is this more noteworthy than the state of Texas, which houses OpenAI's flagship Stargate campus.

Republicans are starting to speak out against rapid data center expansion

The state's governor, Greg Abbott, has undergone a particularly significant change in position. Less than a year ago he described Texas as the "epicenter of AI development," but his campaign for re-election now calls for new restrictions including regulation, reduced tax incentives, restrictions in rural communities and resource consumption measures to protect household bills.

With local opposition so poignant amid ongoing data center expansions globally, Abbott's Democratic counterpart, Gina Hinojosa, also described data centers as "the fight of this election season," making them much more of a political weight than early investors might previously have expected.

Opposition is being driven largely by concerns about the local costs – not an opposition to AI itself. Americans have already seen rising electricity bills, strained water supplies and light and noise pollution.

Ultimately, Texas has shown us how concerns over power prices, water consumption, tax breaks and more can turn technology companies and their projects into serious political cases. And looking ahead, with backlash now spanning both major parties, it's clear that technology companies will need to do more than offer employment opportunities to win over local communities.

Via Reuters

• Lenovo showed Project Swan at IFA 2026, a concept notebook whose 14-inch display rolls out horizontally to a 17-inch 16:9 panel, at 17.9mm and 1.6kg
• It is 90g lighter and 2mm thinner than the vertical ThinkBook Plus Gen 6 Rollable that Lenovo already sells, pointing to a maturing lineup from the computer giant
• It has a compelling business case for people who currently carry a portable monitor for work, but given the current model's pricing, it may come at a significant premium versus the solution it competes with

Lenovo showcased Project Swan at Innovation World during IFA 2026, a proof-of-concept notebook that functions as a 14-inch machine in its compact form and expands into a larger 17-inch 16:9 display when more space is needed.

The laptop measures 17.9mm thick and weighs 1.6kg, and targets customers who do a lot of multitasking, side-by-side document review, presentations, dashboards, and creative workflows.

The model currently has no price tag, no ship date, and no named processor, beyond the company's statement that both concepts, including the fanless AeroBlade, will be powered by Intel at launch this year.

Pitched at users leveraging a portable monitor with a large budget in tow?

Lenovo's Project Swan isn't exactly a new product offering, but an improvement on something the same company already offers. Lenovo already sells its sibling, the ThinkBook Plus Gen 6 Rollable, which went on sale in June 2025 and does the same trick vertically by expanding a 14-inch panel to 16.7 inches, costing at least $3,300 at launch.

Comparing that relatively astronomical price tag to a traditional 14-inch laptop that could cost as little as $1000 from the same company and a portable monitor starting at $200 makes for a very daunting comparison, especially for business buyers, where it could make the difference between a fleet purchase and a niche buy for executives.

It does have some upgrades in thickness and weight over the Gen 6, however: the Gen 6 Rollable is 1.69kg in a chassis 19.9mm thick, while Project Swan is 1.6kg at 17.9mm, shaving off both 90 grams of weight and 2mm of thickness, signaling a maturing technology versus an actual ground-breaking product.

The larger screen and the mechanism to roll it does come with a penalty, however; the ThinkBook 14 Gen 9 announced in parallel with it is 1.3kg at 17.7mm of thickness, beating it soundly in the former and slightly in the latter metric. Against Project Aeroblade, which clocks in at 830g, the weight difference is nearly twice that of the fanless design that Lenovo can currently produce.

The people who benefit are those who need two documents open simultaneously and cannot rely on a second screen being where they are going: contract reviewers, auditors on client sites, translators, analysts comparing a model against its source, consultants presenting across a table without a projector.

For that target audience, the alternative today is carrying a portable USB-C monitor, and the weight calculus favors Project Swan tremendously: A 15-inch portable monitor typically runs 700 to 900 grams, and a comparable 1.3kg ThinkBook plus a screen is around 2.1kg in the bag against the Swan's 1.6kg, with no secondary cable to manage.

Whether Project Swan arrives at an acceptable price point determines whether the rollable laptop becomes more than a pricey proof of concept or a viable solution for many users who need a large screen with portability and weight constraints in tow. For now, Lenovo's impressive offering points to an evolution, not a revolution, in a segment that has drawn a lot of interest but little corporate purchase intent because of concerns about price, durability, and repairability, making it a fringe purchase at best for most.

Following the aftermath of the Manchester Airports Group cyberattack - where hackers made off with the data of 8.7 million people - the hackers have now posted the entire database for sale on the dark web.

The group behind the attack, FulcrumSec, attempted to get Manchester Airports Group (MAG) to pay for the security of the database, promising that it wouldn’t be released if the company paid a settlement. But as all companies should do when faced with extortion, MAG didn’t play ball.

Now FulcrumSec wants to try and maximise the damage of the cyberattack, and has listed the database containing email addresses, phone numbers, vehicle registrations and postcodes online in the hopes a fellow hacking group will find value in the data.

What happens when extortion fails, and why didn’t MAG pay?

In the past, when companies faced ransomware attacks or data breaches, they would sometimes quietly pay the hackers for their silence.

Companies feared serious reputational harm and loss of business would cost more in the long run than the perpetrators were asking for. But this created an incentive for hackers to carry out more attacks.

After all, if companies aren’t kicking up a fuss about being hacked or reporting the attack to the authorities, hacking groups can launch more attacks on other companies that are completely unaware of their tactics.

In order to counter this, authorities and cyber experts from around the world issued guidance that under no circumstances should companies pay for the safe return of their data. By removing the incentive from ransomware attacks and data breaches, the hope is that hackers will get bored or not see enough ROI, and therefore stop.

But the 500GB database of data on over 8.7 million people is ripe pickings for other cybercriminals looking to launch highly specific phishing attacks or scam campaigns. While FulcrumSec may not make any money from the attack, they can seriously enhance the damage of the attack by offering the data for free for others to use.

A successful scam or phishing attack can steal banking and financial details, allowing other groups to steal even more money off the back of the attack.

Expert perspectives on MAG extortion and customer data

• Dray Agha, senior manager of security operations at Huntress:

While Manchester Airports Group followed official guidance by refusing to pay the ransom, the release of 8.7 million records creates an immediate risk for passengers.

We expect other criminals to use this freely available database of vehicle registrations, postcodes and contact details to craft highly convincing phishing attacks.

Anyone who has used parking, lounges or Wi-Fi at these airports must treat unexpected messages about their travel with extreme caution.

• Muhammad Yahya Patel, vCISO and cybersecurity advisor for EMEA at Huntress

The "free release" model is deliberately designed to maximise harm and reputational damage as a warning to the next target.

Publishing almost nine million records for free isn't just punishment for MAG it's a marketing campaign aimed at every other organisation watching.

Pay up, or your customers' data gets handed to every fraudster and scammer on the internet at no cost. Refusing to pay a ransom is the right call. But nearly nine million people are now paying a different price for a decision that was never theirs to make.

• Danny Jenkins, Co Founder & CEO at ThreatLocker

Unfortunately, once data is made public, it can’t be hidden again. The most important thing consumers can do is focus on basic cyber hygiene. Use a unique password for every website, learn how to identify phishing scams, and monitor your credit report.

Be highly suspicious of emails that create a sense of urgency, offer something that seems too good to be true, or ask you to reset your password.

When in doubt, look up the phone number for the alleged sender yourself, rather than using a number provided in the email, and call to confirm whether the communication is genuine.

• Brian Higgins, security specialist at Comparitech:

Whilst the airports breached in this attack don't appear to have had any financial data compromised the risk to affected or associated customers is very real. With so much other information freely available in the wild it is vital that airport users stay highly vigilant for some time to come.

Any and all unsolicited contact; whether online, by telephone or even home visit approaches, should be viewed as suspicious. Never engage until you've taken time to check credentials/veracity etc.

Research advice from trusted sources like the NCA or Information Commissioner and share it with your family and friends. Look for ways to increase digital and physical security like two factor authentication on Apps and devices or Smart Home tech.

When this breach was first reported by the Manchester Evening News the comments were quite telling. Affected parties were quick to identify potential vulnerabilities over and above the breach of financial and banking details.

Home addresses, vehicle registrations and time spent away from home all add up to some excellent opportunities for criminal exploitation, not to mention the usual follow-up phishing campaigns common in this type of incident.

As AI makes data aggregation swift and easy consumers are waking up to the fact that criminals can monetise successful breaches in increasingly inventive ways. It’s no longer enough for data owning organisations to advise post-attack vigilance and turn to their backups.

Victim communities rightly expect better protected networks and systems over and above established norms. As the marketplace grows less fearful and more angry when breaches are made public we may see more emphasis on cyber crime prevention which can only be a good thing.

• Denis Calderone, CTO at Suzu Labs:

The 8.7 million number is attention-grabbing, but it deserves some context. MAG has confirmed that the vast majority of those records are email addresses collected through airport WiFi sign-ups.

A much smaller subset includes phone numbers, vehicle registrations, and postcodes from customers who actually completed parking or lounge bookings.

No payment data, no passwords, no passport information. So despite this affecting airports, which is obviously a sensitive subject, the actual data sensitivity for most affected individuals is relatively low.

What's more interesting to us from a technical standpoint is the attack path. MAG told The Register that attackers compromised one of their internal systems and then went on to steal files from a database hosted by a third party. That's a pivot upstream into a data provider, not downstream into operational systems.

What remains unclear is whether the data was exfiltrated directly from that third-party environment or whether it was pulled back through MAG's network first. That distinction matters for understanding where detection controls failed and who was responsible for monitoring the egress.

The UK's Civil Aviation Authority has a Cyber Assessment Framework for Aviation, developed with the NCSC, that mandates strict separation between IT systems and operational technology.

We don't know whether MAG was formally operating under that framework at the time of this incident, but I would be very interested to find out. Because the segmentation appears to have held here. Flight operations, baggage handling, terminal systems, etc., all were unaffected.

The lateral movement went upstream toward a data provider, not downstream toward the systems that keep planes in the air.

• Seemant Sehgal, CEO and Founder at BreachLock:

This data was initially collected because passengers needed a login, and somewhere along the way, the sensitivity of what was accumulating in that database stopped getting the same scrutiny as the network itself.

Vehicle registration details, postcodes, and contact information across three major airports are a profiling dataset, and whoever held it for ransom understood its value better than the organization storing it did.

• Cisco patched eight IOS XR flaws, including three critical (CVE‑2026‑20274, CVE‑2026‑20279, CVE‑2026‑20212)
• Vulnerabilities allow unauthenticated exploitation, improper access control, and crafted input execution
• No abuse reported; patches urged, with iACL workarounds for Nexus 9000 devices using Silicon One ASIC

Cisco patched eight vulnerabilities affecting its IOS XR operating system, including three critical-severity ones. It urged its customers to apply the patches as soon as possible, even though it stressed that there is no evidence any of these were abused in the wild.

The company detailed its findings in two advisories published on the same day - September 2.

In the first one, it disclosed seven vulnerabilities, including two critical-severity ones: CVE-2026-20274 and CVE-2026-20279. Both carry a severity rating of 9.8/10 (critical). The former is an improper control of a resource during its lifetime flaw - a network-based, low complexity, vulnerability that requires no authentication or user interaction to be exploited. The latter is described as an improper access control vulnerability that can lead to the same consequences.

Fixes and mitigations

These flaws, along with five others, affect all releases of Cisco IOS XR Software, including Cisco IOS XR7 (LNT) Software, regardless of device configuration, the company explained. There are no available workarounds, and installing the provided patch is the only way to mitigate the risk.

The third flaw, disclosed in a separate advisory, is tracked as CVE-2026-20212. Successfully exploiting this one allows attackers to connect to an affected device and send crafted input that could be executed as code, without root privileges. “The exploitation of this vulnerability could also cause the S1HAL process to crash, which could cause the device to reload,” Cisco explained.

This bug affects Cisco Nexus 9000 Series Switches if they include a Silicon One ASIC, the company stressed. A possible workaround is to use infrastructure access control lists (iACLs) to allow only required management and control plane traffic that is destined to the affected device. There is also the option of iACLs only being used to explicitly deny all TCP packets that are destined to a locally configured IP address with a destination port of 43210 or 43211.

Via The Register

• GPT-6 Astra just scored 92.7% on understanding how to interact with your apps, up from 5.6's 76.9%
• New recurrent depth reasoning also allows Astra to think about your problem in a new way
• OpenAI admits there are some safety concerns about how GPT-6 reasons

OpenAI has lifted the wraps off its next-generation model, GPT-6 Astra, and it's being positioned as much more than a conventional chatbot. With this new generation, OpenAI wants us to see its models as being much more suited to computer use.

Being able to fill out forms, update CRM records and organize calendars are among the examples cited by OpenAI, which is especially noteworthy for business customers because it positions Astra as an overarching model that's platform agnostic.

In other words, businesses won't necessarily need bespoke AI APIs or connectors to automate parts of their workflows, because Astra's computer use skill should be able to handle many tasks natively.

GPT-6 Astra could be OpenAI's best business model yet

With its new model, OpenAI is claiming a substantial speed improvement. On the OSWorld 2.0 benchmark, GPT-6 scored 72.6% compared with GPT-5.6's 65.7% and Claude Opus 5's 70.2%. The company's latency simulations put the average task at roughly 40 minutes instead of 75 minutes, marking a near-50% reduction in time.

Astra also scored an impressive 92.7% on ScreenSpot-Pro, which tests a model's ability to understand where to interact with interfaces. GPT-5.6 Sol scored 76.9%, and Claude Fable 5 scored 87.3%.

Among the headline use cases cited by OpenAI in its release information included Astra being able to fully interact with and manipulate Excel spreadsheets, rather than just generating formulas and dumb CSVs, as well as working directly within Power BI to build and refine dashboards, significantly cutting the time a worker could have to spend turning raw data into a presentable format.

GPT-6 also introduces a new reasoning architecture called recurrent depth which allows it to think about a problem repeatedly before producing an answer, which could improve accuracy in high-stakes situations for enterprises.

However, OpenAI admits that monitoring Astra's reasoning has been harder than monitoring Sol's conventional chain-of-thought reasoning, raising some safety concerns, so it's only being rolled out gradually and monitored for now.

Astra is rolling out to Plus, Pro, Business and Enterprise users now, with a Pro variant available to Pro and above. For now, it looks like Free and Go subscribers aren't getting access to the flagship model.

• OpenAI launched Daybreak for Frontline Defenders, offering $1B in credits for resource‑constrained security orgs
• Priority goes to essential services, governments, banks, nonprofits, and open‑source maintainers
• Initiative includes training, partnerships, and a water‑sector pilot with MS‑ISCA for local defenders

OpenAI is offering a billion dollars in credits to security organizations who want to use its Daybreak initiative but cannot afford it.

Daybreak is a cybersecurity program that provides vetted defenders with specialized AI tools and models. It was introduced in May this year, and later expanded into two tiers - Blue, and Red. Blue is the entry-level tier, providing a general-purpose model with custom-tailored safeguards. Red, on the other hand, offers a more purpose-trained cybersecurity model and almost no safeguards.

According to OpenAI, “thousands of defenders across 2,000 approved organizations and workspaces” already use Daybreak, including cybersecurity companies, defense organizations, and law enforcement agencies.

Who gets priority?

But these models cost money, and to make sure more businesses can access them, OpenAI announced, “Daybreak for Frontline Defenders”, a project where defenders can apply and receive credits for AI tokens. Even though the company announced it as a global initiative, it seems that US businesses will be the first to take advantage of it:

“OpenAI is committing $1 billion in subsidized Daybreak access to help resource-constrained cyber defenders, starting with the United States, put frontier AI to work, targeting it to be consumed over the next six months,” the company said in an announcement blog.

Businesses operating essential services such as water and wastewater systems, and electric grid operators, will get priority, OpenAI said. Then come state and local governments, community and regional banks, nonprofits, open-source maintainers, and then “other organizations with limited security resources.”

The subsidy is not just about using the models, though. OpenAI also said it will be increasing hands-on support for frontline defenders, training, and new partnerships.

On top of all that, OpenAI also announced a public sector, water-focused pilot with the Multi-State Information Sharing and Analysis Center (MS-ISCA) to train and support local, tribal, and territorial defenders.

“The pilot will pair Daybreak access with guided training and hands-on assistance for an initial group of public sector and water system defenders, helping them validate and prioritize findings, coordinate remediation, and develop a repeatable approach that can be expanded over time,” the announcement concluded.

• Gemini Audio models power the latest conversational AI in Google Workspace
• Gmail Live can find information buried inside your emails for you
• Users can create Docs and Keep notes with just a brain dump

Google has finally launched the conversational voice features it previewed earlier this May, allowing users to interact with Workspace apps like Gmail and Google Docs by speaking instead of typing.

Much more than just simple dictation, the Gemini Audio models should be able to interpret user intent and retrieve the relevant information.

From launch, the company is supporting its new 'Live' functionality across Gmail, Docs and Keep, but we can expect a fuller rollout as Google continues to tweak the technology based on user feedback and real-world use cases.

Gemini now handles some of your documents and work by voice

Starting with Gmail Live, users can ask things like what gate a flight departs from or what's happening at a child's school during the week, and the model should be able to read and retrieve the necessary information without the user ever having to find the email.

For Docs Live, it's more about generating documents through natural language prompts. Google gave an example of a user giving a 'brain dump' to Gemini, which then goes on to organize the thoughts and produce a structured draft. The tool can also draw information from other Workspace apps like Google Drive, Gmail and the internet.

Keep Live offers similar functionality, turning speech into organized notes rather than simply transcribing a user's thoughts word-for-word.

Google AI Plus, Pro and Ultra subscribers get first access to Gmail Live and Keep Live, beginning now. Docs Live is limited to Pro and Ultra subscribers, cutting access for Plus members. Business customers are also set to get the updates soon.

"Whether searching your inbox, organizing your thoughts, or brainstorming, the new capabilities in Gmail, Docs, and Keep let you do it all conversationally," Google Workspace Product VP Yulie Kwon Kim summarized.

• Hostinger has upgraded its Kodee AI agent
• The tool aims to solve AI fragmentation for small businesses
• It will now reach beyond support, allowing users to create, plan, automate, and analyze work

Hostinger’s AI agent Kodee reportedly solves 91% of 1.5 million support conversations a month, all on its own – a feat so impressive that Hostinger has now given it a new role.

Released earlier this week as Hostinger Agent, the AI agent will now reach beyond support, combining specialized agents that can create, plan, automate, and analyze work.

Hostinger Agent retains the Kodee approach, aiming to solve as many tasks and issues on its own, but handing cases off to human specialists when it is unable to do the job.

AI fragmentation for small businesses

According to a press release shared with TechRadar Pro earlier this week, Hostinger Agent aims to solve the issues of AI fragmentation for small businesses.

Today, many small businesses deal with bots for every vendor, plus their own SEO tool, writing copilot, and analytics agent. Each requires time to set up and get familiar with. Plus, they have no knowledge of the others, often leading to repetitive conversations.

Research by Gartner predicts that 60% of brands will be using agentic AI across key business areas, including marketing, sales, and support, by 2028 - highlighting just how AI fragmentation could compound in the coming years.

“The AI industry has spent the past two years giving people more assistants, copilots, and agents to choose from. We want customers to stop choosing between them,” said Emilis Strimaitis, Head of Product Innovation at Hostinger.

“When the AI already knows what the customer is building and what they are trying to achieve, fixing the immediate problem is only part of the job. The real value is helping them grow online.”

From advice to action

Hostinger Agent solves this by connecting support, marketing, SEO, content and visuals, analysis, and recurring tasks into one interface, letting users launch and manage their website in one space.

The AI agent doesn’t just offer advice; it takes action across your website and within hPanel, reducing the technical know-how and time typically required. By consolidating all of these into one place, it will also be able to develop a better contextual understanding of your business, supporting better advice.

Users will also be able to access Hostinger Agent within AI Builder, Hostinger’s dedicated space for creating websites, stores, apps, and other online projects with AI agents, opening the door for users to build websites and consult with the AI agent at the same time.

Chips from Taiwan, oil from the Persian Gulf, and cars from Germany – what was long taken for granted is now being called into question all over the world. Where markets and expertise are concentrated among a few providers, enormous economies of scale arise.

At the same time, however, dependencies also grow – often with risks that only become apparent when supplies fail, prices rise, demand evolves, policies change, or individual players begin to monopolize the market.

Take AI, for example – according to research institute, Epoch AI, only two notable AI models have emerged in Europe over the past twenty years compared to 50 in the US and 30 in China. The technology has moved so quickly that the repercussions of this imbalance are only just beginning to be felt in European markets.

Is Europe becoming too dependent?

From office applications to storage and computing resources to software, what providers deliver from their data centers in a highly scalable, maximally available, and extremely innovative manner does not, in most cases, originate from the European continent.

According to Bitkom’s Cloud Report 2026, roughly 85% of German companies consider Germany to be too dependent on US providers, and almost two thirds of local users (64%) feel compelled to review their own IT strategy due to US government policy. Countries like France are already taking action, with the French government announcing plans to replace Windows with Linux on its government computers.

At the same time, the national health insurance fund has already migrated 80,000 employees from Microsoft Teams, Zoom, and Dropbox to its own alternatives.

Whether in Germany or France, the question of how digitally dependent companies, governments, and societies are on individual suppliers is being asked throughout Europe. And it has been for some time. While Gaia-X focused on common standards for sovereign cloud computing, the EU is now pursuing a more pragmatic approach.

The Cloud Sovereignty Framework (CSF), unveiled in October 2025, aims to make digital sovereignty measurable. To achieve this, it builds on initiatives such as Gaia-X as well as regulations like DORA and NIS2, with the ultimate goal of being able to prioritize compliant, low-risk providers in public procurement.

The sovereignty metric score

Controlling and regionally consolidating one’s own digital traffic flows. That’s quite a task in a world where AI superpowers are evolving overseas. With regard to IT supply chains, the framework favors providers who operate their environments in a transparent, resilient, and sufficiently sovereign manner to ensure that systems are protected from external influences or the interests of third parties.

This means, for example, that data paths must be traceable and controllable at the regional level to ensure they remain within the EU legal framework. In other words, there is no sovereign cloud without an interconnection strategy. After all, Internet Exchanges (IXs) already play a key role when it comes to processing bits and bytes locally. Wherever companies, networks, and clouds connect at IXs, they exchange information along paths that can be precisely defined geographically.

Although the EU does not explicitly make IXs a scoring criterion, it’s heavily implied. For example, the CSF also evaluates cloud services based on supply chain-related legal, operational, and technological objectives. This concept is also reflected in the NIS2 Directive and argues in favor of more decentralized and resilient backbones, such as mutually secured interconnection platforms, such as IXs.

In practical terms, this means that any IT landscape is only as secure as the individual elements that comprise it. So, if all partners design the components they provide to others with multiple layers of redundancy, the overall system becomes not only more reliable but also more independent for everyone.

Distributed and provider-neutral IXs have already made this principle their guiding philosophy: to allow data to be exchanged with high availability, security, and control – between all companies, partners, and clouds within the digital ecosystem.

Europe at a crossroads

Technical issues can usually be resolved easily, but legal ones often cannot. Therefore, the CSF is unlikely to revolutionize the European cloud market overnight, but it can steer public demand in a targeted manner. As a result, hybrid, multi-cloud, and on-premises approaches are becoming more important, and IT is becoming infinitely more complex.

While unravelling this complexity will take additional effort, the pay-off of rearchitecting networks around neutral IXs will be a two for one: not only will it reduce long-term lock-in risks but will also reduce the threat of having to pay significantly higher prices later on when things inevitably change. Europe must now consciously accept this and align and strengthen its own provider market accordingly.

Just as operations would come to a standstill for 46% of companies in a prolonged outage, without the cloud, there can be no functional public administration – and no independent, sovereign, and thus capable state.

We've featured the best cloud storage.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

When businesses and governments delegate operational tasks to AI agents, the discussion tends to start with capability. Can AI do what a person can do, such as reconcile accounts, negotiate with suppliers, sign a document or submit a public filing? And can it do so quickly and cheaply enough to justify putting it to work?

These questions matter, but performance is only part of the picture. Once an agent acts independently, different questions emerge; who authorized it, whose interests does it represent, where does its authority end and, perhaps most importantly, who is responsible when something goes wrong?

This is the accountability gap in agentic AI, and a more advanced model will not close it. An agent might be remarkably capable, yet still act with potentially unwanted or possibly even unlawful consequences.

The only way to bridge that gap is through a system with clear permission guardrails and an audit trail that links each agentic action to the person that authorized it, and can tell the difference between a person and a machine in the first place.

Ordinary error at extraordinary scale

When people think about AI risk, they tend to picture a machine breaking free of human control and pursuing goals of its own. The more immediate risk is less dramatic; a system simply getting an ordinary task wrong, again and again, with no clear line of responsibility and accountability.

A person might send an invoice to the wrong customer. An agent connected to the full customer database could send the same invoice to thousands. It might change the wrong field and then repeat the change across an entire market, not because it has become malicious, but because it has misunderstood an instruction and had the access to act on that.

The speed that makes an agent valuable can save a business hundreds of hours, but can also turn a small and recoverable mistake into a serious operational problem.

When an agent drafts an email for someone to review, this ambiguity may not matter very much because the person remains between the software and the outside world.

But when it can send the message, submit a declaration or move money, it matters a great deal. Imagine asking an assistant to book one flight and, instead of giving them a company card with an appropriate spending limit, handing over your passport, bank card and office keys. You may trust the assistant completely, but the permissions bear no sensible relation to the task.

Define the boundaries, solve the problem

AI agents cannot hold citizenship or legal personhood, but they can be linked to natural persons in legally sound ways. For example, a reliable identifier could trace the agent's activity back to the person responsible for deploying it. Combined with clear permissions and mandates that make the terms of its authority visible and auditable, this would allow AI agents to carry out a range of operations and transactions in a more controlled way.

These controls could be very precise. An agent might be allowed to view financial information but not change it, prepare a payment but not approve it or order from named suppliers without exceeding a fixed amount. Its permissions might last for one transaction, one working day or the duration of a particular contract, after which it would expire without relying on someone to remember to switch it off.

Withdrawing that authority should be equally straightforward. If a company ends its relationship with an agent, changes supplier, or discovers that something is wrong, it should be able to cancel the it’s access without changing the credentials of the person behind it.

The agent’s legally anchored mandate in the various realms of these controls has a longer shelf life than any individual AI model, which will be upgraded, replaced and combined over time, often without the customer seeing much more than a new version number.

Putting the idea into practice

Estonia has decided to work through these questions now. In June, the government began developing a state-backed registration system for AI agents that will be available to both Estonian citizens and the country’s global community of entrepreneurs. The goal is to explore how Estonia’s fully digital state, where all public services are available online, can be opened to the emerging wave of AI-powered businesses.

Under the current proposal, a person would receive a registered numeric identifier that could be linked to one or more agents acting on their behalf. Operations performed by an AI agent are treated as operations performed by a machine, for which the natural person, the identity anchor, is liable. The AI agent can only operate within the limits of the authorizations granted to it.

For transparency, private service providers also need to know whether an operation was performed by a person or by a machine acting on their behalf. This allows them to apply appropriate controls, restrictions and risk models.

Built for delegation

The approach draws on more than two decades of digital government. Estonia’s system of personal identifiers and digital identities already allows people to act on behalf of others with clearly defined permissions.

An accountant can file a client’s taxes, an adult can manage an elderly parent’s affairs through the health portal, and several people can use a corporate bank account with individual rights and limits. In principle, the same framework could support mandates for AI agents.

This would give users greater confidence that safeguards are in place to protect the responsible person from foreseeable errors by an AI agent. It would also help close the accountability gap as agentic activity increases.

Before AI agents are given executive powers, they need a technically and legally binding framework that records who they represent, what they can do and who remains responsible for their actions.

We've featured the best AI tool.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Tottenham Hotspur just confirmed plans to migrate from VMware to HPE, saving 85% off its bill
• CTO Rob Pickering revealed that Broadcom-era changes are largely to blame
• The stadium is also undergoing a broader HPE upgrade

Tottenham Hotspur has phased out VMware at its stadium as part of a major infrastructure overhaul. By moving to HPE Morpheus VM Essentials, the London football team managed to cut virtualization costs by more than 85%.

According to Ars Technica, CTO Rob Pickering explicitly pointed toward the commercial changes following Broadcom's acquisition of VMware.

But for Tottenham, the migration was just one part of a major data center and networking refresh, with the company also shifting its server, storage and networking infrastructure – incidentally, to HPE.

Tottenham moves from VMware to HPE, saves 85%

Though the team didn't reveal how much it paid for which products with VMware, Pickering did say the saving was a "significant amount" and "north of 85%."

"The issues with the Broadcom takeover of VMware are well-known in terms of the commercial changes to their structure and the way that they engage with clients," he added.

Tottenham also confirmed that integrating virtualization, infrastructure and automation into one single operational stack ultimately led to more value than simply swapping one virtualization vendor for another to save money, hence the decision to go all-in on HPE.

Earlier this summer, HPE announced that it would be offering prospective customers a year's free HPE Morpheus VM Essentials for customers switching from VMware to help ease the cost involved with migration. Its Zerto migration software was also confirmed to be part of a $1 package.

It's unclear whether Tottenham Hotspur took advantage of those offers.

At the same time, the stadium is also said to have started a major network upgrade toward Wi-Fi 7 and HPE Aruba Central in a stadium with around 20,000 network access points and connections.

More broadly, Tottenham Hotspur isn't the only business turning its back on VMware – UK supermarket giant Tesco also confirmed plans to dump 40,000 VMware licenses.

Agentic commerce has cleared its first hurdle, that of consumer demand. ChatGPT handles an estimated 50m shopping queries a day, and AI-referred traffic to US retail sites rose by nearly 400% in the first quarter of 2026. The unresolved question is where that demand settles once a purchase is made. The early evidence points to the scaled, branded marketplace.

According to Similarweb, marketplaces already receive more AI-referred traffic than any other retail category, some 47m visits in the year to May 2026, and continue to grow faster than their scale would predict.

The mechanism is a mutually reinforcing relationship between consumer preference and LLM recommendation, and its logic points towards a concentration of agent-led commerce on large platforms, with material consequences for merchants and their payment suppliers.

The pattern in AI referral traffic

Revealed consumer preference is a much more reliable guide to market development than stated intent, and the traffic data for agentic commerce is gradually building a picture of how consumers actually use agentic commerce rather than just speculation. Similarweb’s recent analysis ranks marketplaces first among retail categories for AI-referred visits, ahead of news, travel and finance.

Two features are notable. The category commands the largest absolute volume of click-through traffic, and it continues to expand at more than 230% a year despite that already substantial base. The explanation lies in how large language models (LLMs) assemble their recommended shortlists.

Faced with an open-ended shopping request, a model favors platforms with broad catalogues, well-structured product data and extensive reviews. Large marketplaces can provide these in abundance by providing a single source that can satisfy the widest range of queries with the strongest supporting evidence.

Testing the democratization thesis

A contrary hypothesis has been influential. Shopify’s president, Harley Finkelstein, has characterized agentic commerce as ‘fundamentally merit-based’, arguing that it would widen discovery to the long tail of smaller merchants rather than reinforce the incumbents that dominate conventional search.

The demand is not in doubt: Shopify reports that AI-referred traffic to its stores has risen roughly eightfold in a year. The distribution of that demand, however, is less even than the thesis implies.

Marketplaces, not independent sellers, are absorbing the largest share of AI referrals, and the agentic platforms attempting broad merchant onboarding have found it difficult. OpenAI has scaled back open instant checkout in favor of a smaller set of curated partners, citing the complexity of onboarding and verifying merchants at scale.

The emerging structure is therefore symbiotic. Marketplaces supply a large, trusted brand within which a long tail of smaller sellers operate. The individual merchant still reaches the agent-guided shopper, but within a brand and trust environment the consumer and the model already recognize.

This arrangement addresses the consumer trust problem that the open web does not, since the platform has verified the seller before the listing appears, and it retains the operational layer, fulfilment, dispute resolution and logistics, that an agent appear unwilling to take responsibility for. The long tail is not excluded so much as commercially enabled by marketplaces.

A self-reinforcing loop

Beneath the traffic figures lies another dynamic that has received relatively little analysis. Consumer preference and algorithmic recommendation reinforce one another. Consumers already trust marketplaces in a traditional ecommerce use case and select them when an agent presents the option.

Learning models, optimized on the recommendations consumers select, are more likely to surface marketplaces more frequently, reinforcing the habit and shaping the next recommendation. Over successive interactions the marketplace becomes both the more probable choice for the shopper and the more probable output of the model.

New international survey of agentic shoppers indicates that these outcomes are already in consumers’ minds. Consumers tell us they are looking for an independent (i.e. non-seller embedded) agent to compile the shortlist, followed by the reassurance of a familiar brand and visible reviews, a combination that large LLM brands and marketplaces are well placed to provide.

The research also showed that 74% of agentic shoppers in the US and Europe favor an independent assistant capable of comparing across sellers, against 10% who prefer one embedded within a single seller. Brand recognition is also important to 89% of respondents, and customer reviews to 93% when deciding which recommendation to pick from an agent’s shortlist. Most significantly, 93% expect to use retail marketplaces as much as or more than before as the adoption of agentic commerce grows.

Why marketplaces hold the advantage

Consumer preference is only part of the explanation. The economics facing the agent point in the same direction, for three reasons.

The first is remuneration. Most large marketplaces operate established affiliate and commission schemes, providing a ready mechanism to reward an agent for delivering a customer. As agents assume the role once played by referral sites and publishers, the marketplace platforms with mature payout infrastructure are the simplest to monetize.

The second is conversion. Marketplaces have invested heavily in low-friction checkout, so an agent compensated on completed sales rather than referrals has reason to direct shoppers where a purchase is most likely to conclude, which compounds the commercial pull towards the platform.

The third is payment. Agentic commerce already leans heavily on credentials held on file. The dominant standard, Google’s Universal Commerce Protocol (UCP), only transacts against tokenized credentials held on file rather than card details entered by hand, and marketplaces already hold card credentials for hundreds of millions of shoppers.

The advantage is not only the agent’s. Agentic commerce is still nascent: technical standards are still emerging, liability and commercial terms are unresolved, and fraud is migrating to the channel. Absorbing that uncertainty requires capital, specialist staff and an appetite for risk.

Large marketplaces can invest through the ambiguity, adopt and switch between emerging protocols, and operate the fraud and dispute controls the channel demands. For most mid-sized and smaller merchants, the same burden is disproportionate to their means and their tolerance for risk, a further reason to participate through a platform rather than to carry the complexity alone.

An acceleration, not a departure

The movement towards marketplaces predates agentic commerce, and there is a long list of large retailers that have already built or joined them. Kingfisher opened a third-party marketplace at B&Q in 2022, and such sales now represent around 40% of B&Q’s online revenues.

Tesco launched Tesco Marketplace in June 2024 and listed more than 300,000 third-party lines within eight months. Target has told investors it intends to expand its invitation-only Target Plus marketplace from more than $1bn to more than $5bn in gross merchandise value within five years. In each case a trusted retail brand hosts a long tail of smaller sellers, the configuration the referral data now rewards.

Implications for merchants and acquirers

The likely consequence is that agentic commerce accelerates this migration, and that the reinforcing loop strengthens rather than dilutes the move towards marketplaces. The effect will likely divide according to the scale and strength of a merchant’s brand.

Larger, well-known retailers may choose to become marketplaces in their own right, hosting third-party sellers beneath their own name, as Kingfisher, Tesco and Target have done.

For smaller merchants the choice is more likely to be additive: most will want to retain a direct presence while also selling through one or more established banners, drawing on the reputation of the host and a position within the recommendation loop.

This is the next stage of a long-standing movement of smaller firms into marketplace distribution rather than a break from it.

Direct commerce will not disappear. Merchants with strong brands, machine-readable reviews and genuine loyalty can still secure the agent’s shortlist independently, and Salesforce reports that retailers operating their own shopping agents have grown sales roughly 60% faster than those without.

For the broad middle of the market, however, and particularly in impersonation-prone categories such as fashion and footwear where consumers and their agents may have concerns about seller trust, a marketplace presence may become a competitive requirement.

For acquirers, the shift accelerates a migration of volume that is already underway by moving spend onto a smaller number of large platforms. Meanwhile the direct-merchant portfolio shrinks and may be exposed to elevated fraud risk as fraudsters exploit the weakest validation layer.

Acquirers with strong exposure to the marketplace segment are well placed, while the rest of the market faces yet another reason to move into the platform economy.

We've featured the best ecommerce hosting.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Most manufacturers didn't need a business case to understand the cost of reactive maintenance. They'd been absorbing it for years – unplanned stops, recovery overtime, expedited parts, the slow erosion of scheduling confidence and customer trust. What shifted was the available answer. AI in maintenance moved from speculative to deployable fast enough that the investment case made itself.

Industry commitment is real, but something isn't converting at the rate anyone projected. Our recent research found that approximately 78% of all reported barriers to progress are workforce-related. Access to AI moved faster than the ability to use it consistently. That gap is now the constraint.

The first wave proved the tools, not the model

Manufacturers are investing in AI to improve productivity, yet many are still carrying the behaviors that predictive maintenance was meant to reduce.

The research shows predictive maintenance adoption has more than doubled year over year, while reactive maintenance remained flat. Proactive maintenance has also lost ground year over year. That combination is the point: new methods are entering the plant, but they’re not fully replacing the old ones.

This shouldn’t be read as failure. Plants moved for practical reasons, and early pilots gave teams useful proof. A model can work well on a known asset with a focused team around it. The harder test comes when that model has to support decisions across shifts, sites, and mixed levels of experience.

That’s where the first wave exposed the next problem. Technology can move quickly into the budget. Work habits, trust, decision rights, and frontline confidence take longer to change.

AI investment grows up

The budget data confirms that leaders aren’t walking away from AI; they’re becoming more selective about where it has to prove itself. Investment is moving away from exploratory AI and toward operational priorities, including cybersecurity, data management, Generative AI and Industrial AI.

That shift reflects a more practical view of digital maturity. Leaders are trying to make AI work where the cost of delay, downtime, and poor data shows up quickly.

It also reframes expectations around Industry 5.0. As industrial technology moves from Industry 4.0’s automation-led model toward a more human-in-the-loop approach, leaders appear to be recalibrating the timeline, with 40% now expecting a one- to four-year journey.

Where the model stops and the supervisor starts

It’s easy to misread the 78% workforce-barrier figure as a labor shortage story – headcount, hiring, pipeline. However, the truth of the matter is that businesses are facing a lack of expertise, knowledge shortages, skilled labor gaps, and broader workforce capability deficits. Taken together, those four categories describe something harder to fix than a recruitment problem. They describe an organization's capacity to absorb a different way of working.

It's a pattern I hear consistently from customers: the tools are in place, but the teams around them are still catching up. Researchers Cohen and Levinthal gave that capacity a name: absorptive capacity.

In plain English, it’s how quickly a business can recognize useful new knowledge, absorb it into the organization and convert it into practical output. In a maintenance context, that output is a decision made under real operating conditions – by a night-shift supervisor weighing whether a current anomaly justifies an intervention, whether it can wait until the next planned stop, or whether the risk has already crossed a line. A model can flag the anomaly. It can’t make that call.

What predictive maintenance and AI actually demand from the workforce is harder to train than tool proficiency. The UK Government’s AI Skills for the UK Workforce report identifies the gap in concrete terms for advanced manufacturing.

Beyond technical application skills like training AI models or integrating real-time analytics, the report points to the ability to interpret AI outputs, adapt workflows around new insights, and communicate changes to frontline teams. These are key competencies that determine whether an alert changes behavior or gets dismissed.

That’s why the workforce issue can’t sit beside the technology program as a separate HR workstream. It’s part of the system that decides whether investment turns into better execution. Industry 5.0’s timetable will be set by capability on the plant floor.

Investing for predictive, paying for reactive

The commercial consequence is visible in the lag between spending and return. Predictive adoption is rising, more capital is being allocated, yet the reactive baseline hasn’t moved. Many plants are running two modes of operation at once – investing for data-driven execution while still losing time to avoidable firefighting.

That split has a cost. Siemens’ 2024 True Cost of Downtime research puts unplanned downtime losses for the world’s 500 biggest companies at $1.4 trillion annually – 11% of their total revenues. That figure reflects what happens when the shift from reactive to predictive is incomplete. The investment is there, but the daily execution hasn’t fully followed.

When that gap persists, ROI on technology spend arrives slowly and unevenly. Timelines extend. Confidence in delivery weakens.

Where the return is earned

The discipline being applied to tools and platforms now needs to apply to the people and routines around them. For many organizations, that’s where the return is being left on the table.

Leaders should start with the parts of the operation where capability is most fragile. Which assets still depend on one or two experienced technicians to interpret what’s happening? Which work histories are too thin to support the next diagnosis? Which alerts trigger confident action, and which sit in limbo until the right person is on shift?

Those questions reveal whether predictive maintenance has become part of execution or whether it’s still sitting on top of reactive habits.

Capture the know-how that still lives in people’s heads before it walks out the door. Make work histories complete enough to help the next technician. Train operators and maintenance teams to understand what a predictive alert is telling them, what evidence warrants action and when to escalate. Shape workflows so acting on insight becomes the normal path, not a special effort.

Connected reliability supports that execution when it stays close to the work – connecting asset data, maintenance history and frontline judgment so teams can make repeatable decisions across shifts and sites.

Our research backs this up: nearly half of respondents plan to advance connected reliability initiatives within the next 12 months, treating reliability as the practical bridge between near-term operational needs and longer-term ambitions.

The call for manufacturing leaders is straightforward: audit capability with the same seriousness as technology spend. Don’t stop at asking what AI has been deployed. Ask who can act on it, where decisions slow down, what knowledge is undocumented and which workflows still pull teams back into reactive work.

We've featured the best AI tool.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Four federal agencies criticized for failing to share the framework's details
• We don't know anything about who's involved and how models are tested
• Protect Democracy wants a court order by the end of the month

Four separate US federal agencies are being sued over the Trump administration's secret framework for testing frontier AI models before they get released.

Protect Democracy argues that "almost no details" have been shared about how the framework works, who participates in it, how companies are selected or what legal authority underpins the reviews.

The nonprofit warns the secrecy is especially significant because the government could potentially influence which AI models can be released and who can access them.

US government agencies sued over AI framework

This comes after weeks of battle between Claude-maker Anthropic and the US Government following revisions to the availability of Mythos 5 and an attempt by President Trump to blacklist the company.

Protect Democracy wants the government to disclose the framework's details, its participation terms, the identities of participants, rules governing access to frontier models and the contractual arrangements. The group also wants an injunction preventing agencies from withholding records that are not classified.

The government bodies in question include the Office of the National Cyber Director, the Office of Science and Technology Policy, and the Departments of both the Treasury and Commerce.

The White House has argued that aspects of the evaluation legitimately need to remain a secret, and national security-related tests could contain classified information.

"Just because things are unclassified, that doesn’t mean we are going to broadcast them to everyone," a White House spokesperson said.

Protecting Democracy isn't the only entity concerned about the framework. US Representative Greg Casar (D-Tx.) accused Trump of "completely failing to keep us safe from the dangers of AI" and being "too busy cashing in to protect our jobs or national security."

With Congress also in the dark, Protect Democracy argues the details are needed for both congressional oversight and public accountability.

The group is now seeking a court order for disclosure of the non-classified framework and related details by September 30.

• Chinese companies have spent the past four years filling gaps left by US bans
• The number of IPOs centered around semiconductors has increased
• China is also concentrating on supply chain self-sufficiency

New Morgan Stanley data cited in a South China Morning Post report has demonstrated a shift in the companies reaching China's public markets, per the analysis of 229 IPOs on Shanghai's Star Market between 2022 and 2026.

The number of companies targeting China's technological 'chokepoints' has become much more prominent – around one in five of the IPOs fell into this category in 2026, compared with fewer than one in ten in 2022.

In other words, the SCMP claims that the number of companies addressing areas of foreign technological dependence has more than doubled in just four years.

China's public market fills in response to US restrictions

In the four years analyzed, the so-called chokepoints have concentrated around the semiconductor supply chain following US restrictions and sale bans.

But by preventing the sale of some of its own chips and hardware, the US has fundamentally enabled China to grow its own market by giving it an incentive to develop domestic alternatives.

Morgan Stanley identified 21 companies in 2026 that were specifically targeting these shortfall areas – 19 of them were closely aligns with the semiconductor supply chain.

The investment banking giant also estimated that around 60% of the companies listing in 2026 contribute to China's supply-chain self-sufficiency drive, up from 41% in 2022.

While some restrictions have been eased during that time period, China continues to invest heavily in this market, with continued investments across key remaining areas like raw materials and manufacturing machinery.

Ultimately, years of US technology restrictions have increasingly pushed Chinese startups towards replacing that foreign technology, and even if restrictions were to stop overnight, it's likely that China could now be in a position not to need US hardware anymore anyway.

• Microsoft warns of Teams‑based campaign where attackers impersonate IT staff
• Victims tricked into granting remote access, leading to malware, lateral movement, and ransomware
• Defenses: verify support contacts, train staff, harden Teams, and use Defender Safe Links/ZAP

Microsoft is warning about an ongoing hacking campaign that starts with a Teams message and ends with a ransomware infection and data theft.

In a new in-depth report published on the Microsoft blog, it was said that unnamed threat actors were reaching out to their targets at various enterprises via a Teams chat, while impersonating IT staff.

They were coercing their victims into granting remote access via screen sharing or remote monitoring and management tools and once received, used their access to install malware loaders and various other implants.

How to defend against Teams-borne phishing

The malware was just the first stage of the attack. Subsequent stages include host reconnaissance, security-product and virtualization discovery, and “periodic desktop screen capture”. In other words - mapping out the landscape and conducting espionage.

The crooks would then enumerate domain accounts, servers, and users, through native tools and Active Directory Service Interfaces (ADSI) queries and begin moving laterally.

The final step includes identifying and extracting valuable data, followed by a ransomware infection.

Microsoft does not name the perpetrators, and mostly refers to them as “threat actors”. It makes sense, since the “fake IT support via Teams” technique is being used by multiple groups at this moment. Russia’s Cozy Bear, FIN7, and Storm-1811 are probably the most obvious examples.

The world’s biggest extortionists - ShinyHunters - are also known to use Teams to trick victims into granting access, but this group rarely deploys an encryptor and instead just focuses on data exfiltration.

Whoever the attackers are, and whoever they’re after, one thing is for certain - the risk in the enterprise environment has never been greater.

That is why Microsoft advises reinforcing user education by establishing internal helpdesk authentication phrases, and by training employees to recognize external-tenant indicators.

The company also urges enterprises to verify unsolicited support contact, and to harden Microsoft Teams and email against social engineering. “Use Microsoft Defender for Office 365 with Safe Links and Zero-hour auto purge (ZAP) so malicious messages and URLs are neutralized at time of click and removed after delivery,” Microsoft urges.

• Lenovo showed Project AeroBlade at IFA 2026, a fanless 14-inch concept under 10mm thick and under 830g, cooled by a solid-state Active Flow system built with Frore Systems
• Lenovo dubs this a 'quieter, cleaner computing experience' sans the performance tradeoffs that many similar designs have to make
• Published AirJet specifications and Intel's own 11.3mm reference design put sustained cooling near 15 watts, so "full-performance computing" from Lenovo might be a reach at best currently unless its purported breakthrough offers considerably more cooling than the former's already established tech

Among the shipping products at Lenovo Innovation World during IFA 2026 are two proof-of-concept machines, and the more consequential of the two, somewhat ironically, is not the one with the moving screen.

Project AeroBlade is a 14-inch notebook measuring under 10mm thick and weighing under 830g, built from magnesium-aluminum with a high screen-to-body OLED display, Thunderbolt 4, and Wi-Fi 7 that eliminates fans in the process.

Lenovo cools it with what it calls Active Flow, a solid-state thermal architecture developed with Frore Systems and built on top of the latter's AirJet technology.

Big promises, but relatively scant details

Lenovo claims the resulting product delivers full-performance computing and sustained AI performance without the compromises normally associated with fanless designs.

This could depend on what one considers 'full-performance' computing and AI use, as well as the underlying CPU and GPU configuration. Frore Systems' AirJet is not vaporware; it is proven technology that uses a piezoelectric cooling chip to move air with membranes vibrating above the range of human hearing, and the solution has shipped in mini PCs and SSD enclosures for two years.

An AirJet Mini removes 5.25 watts of heat at 21 dBA while drawing up to one watt, in a package 2.8mm thick and weighing nine grams, generating 1,750 pascals of back pressure. A Mini Slim variant trims that to 2.5mm at the same capacity. The higher-capacity AirJet Pro, rated at 10.5 watts, has since been discontinued, so scaling now happens by adding 5.25-watt modules rather than upgrading to a larger unit.

The tech, however, has not been abandoned or ignored, with Intel building a Wildcat Lake laptop reference design around an AirJet Mini wrapped in a vapor chamber, and Tom's Hardware reporting 15 watts of sustained cooling in a chassis 11.3mm thick.

Lenovo's AeroBlade, however, is thinner at just under 10 mm, which would make it a small achievement if it still sustains 15 W of cooling, but it also means Lenovo is being pickier than most when it comes to choosing facts here: it may just be running a low-powered chip at "full-performance".

For context, the ThinkBook 14x Gen 2 Lenovo is shipping from Q1 2027 carries a 25W TDP CPU under the hood, and running that at maximum design power would require a significantly larger cooling mechanism: the equivalent of five AirJet Minis and a vapor chamber to match.

Lenovo's concept does have advantages, however: it will be much quieter than any design that uses fans, while also removing a key component that can cause hardware failures or RMAs, making longevity a key gain for future designs based on the same approach.

For now, however, Lenovo has not attached a price, a date, a processor, or a single executive quote to the proof-of-concept laptop, which effectively means that it is years from becoming a viable mainstream option. It has, however, demonstrated that a 14-inch business laptop under 830 grams with Thunderbolt and no moving parts can be built with components that exist today, and that alone is arguably a sizeable victory for enthusiasts wanting a quiet but powerful x86-based alternative to the passively cooled MacBook Air.

• Lenovo announced the 11th-generation 15" Yoga Pro 9n and 16" Yoga 9n 2-in-1 at IFA 2026, both running NVIDIA's RTX Spark with a 20-core Grace Arm CPU and a 6,144-core Blackwell GPU
• The understated highlight for business users and professionals is the 128GB of unified memory that the Pro 9n offers, essentially allowing for large LLMs to be run on-device
• This is essential for companies that are required not to transmit sensitive data or use cloud APIs for their AI requirements

Lenovo announced two Yoga laptops running NVIDIA RTX Spark at its Innovation World event during IFA 2026, and the spec that may matter most to professional buyers isn't the GPU but the massive maximum memory both offer.

While the Lenovo Yoga 9n is limited to a 16-channel 256-bit 64GB LPDDR5X configuration, the Pro SKU doubles this to 128 GB.

The change allows the Yoga Pro 9n to run large models on-device, an impressive feat for a device that weighs only 1.65kg but houses a capable Blackwell RTX GPU that can muster as many as 6144 cores for local AI needs.

Peaking under the hood of the RTX Spark-powered Yoga laptops

The underlying silicon is NVIDIA's RTX Spark, unveiled on May 31 ahead of GTC Taipei and long rumored under the codename N1X. Both laptops pair a 20-core Grace-class Arm CPU with a Blackwell RTX GPU with 6,144 CUDA cores, roughly comparable to a laptop with a GeForce RTX 5070, co-developed with MediaTek and joined by Nvidia's proprietary coherent chip-to-chip interconnect (NVLink-C2C).

The Lenovo Yoga Pro 9n is the performance-centric model, offering 2 NVMe SSD slots and a total advertised storage capacity of 4 TB, while the non-Pro model caps at 2 TB. It also offers an extremely color-accurate 15.3-inch 2.5K PureSight Pro OLED that runs from 48 to 165Hz at 1100 nits peak brightness, plus a large 92.5 Whr battery with rapid charging that lets it eke out an advertised 3 hours of battery life after 15 minutes of charging. It also offers 2 USB-A and 2 USB-C (USB4) ports; the latter support DisplayPort 2.1 and USB PD 3.0 for power users, and it also includes an HDMI 2.1 connection and an SD port.

With WiFi 7 and BT 5.4 under the hood, plus a 9.2MP IR camera and a quad-speaker setup, the Pro 9n is built for creators who need plenty of connectivity in a slim, portable chassis.

For those who need a slightly larger screen and potentially a 2-in-device, the Yoga 9n skips the 'Pro' moniker but offers a 2.8K OLED display with a slightly lower capped refresh rate (30 to 120Hz), half the maximum memory configuration, and a slightly larger battery (99.9 WHr), even as it sees its weight increase to 1.94KG as a result.

A portable for-local AI offering that holds its own

For business buyers looking to localize their AI needs, run large proprietary AI models, or keep data on-device, the unified memory pool both Lenovo offerings bring to the table is essentially their biggest selling point.

This is in sharp relief from how the first wave of Copilot+ PCs stressed AI performance as a unit of NPU TOPS, even as the emergent bottleneck was that of memory capacity and bandwidth, where both laptops excel.

A machine that holds a large model in memory can essentially run inference on material that never leaves the device, even as both laptops offer a decent mid-range GPU that can generate more tokens than most comparable laptops for similar models.

This is crucial for firms that cannot send client documents, patient records, case files, source code, or export-controlled technical data to a third-party cloud API; this is not just a performance feature but a compliance one, and it marks the difference between using a model on real work with information that is often sensitive versus on test data.

The laptops also bring relief for developers who spent the last few years trying to use unified memory-based Apple laptops to work with Nvidia's CUDA stack with varying amounts of success, to build locally on a computer that essentially eliminates that compatibility and testing gap by allowing it to run the CUDA stack natively on the RTX Spark-based machines.

One could argue that many applications still have compatibility issues on ARM-based CPUs, but that gap is narrowing as Microsoft continues to publish updates and emulation improves, and as more developers publish ARM-centric compilations of their software.

Despite their AI-centric focus and catering to high-end professional users, Yoga generally isn't where Lenovo's business customers are found; its business fleets run on ThinkPad, with ThinkShield security tooling, business-grade manageability, extended parts availability, and the lifecycle commitments that enable predictable refresh cycles.

Lenovo has not announced pricing for the Yoga Pro 9n or the Yoga 9n currently, stating that it will do so at launch, even as the accompanying Snapdragon 8 Elite-powered Lenovo Yoga Tab Plus Gen 2 and MediaTek Dimensity 9500s-powered Lenovo Yoga Tab Gen 2 both reach shelves this September with price tags of $999.99 and $729.99, respectively.

• Lenovo announced the ThinkBook 14x Gen 2 at IFA 2026: a 990g notebook with up to a 14-inch 2.8K 120Hz OLED option, Intel Core 7 Series 3 silicon, and a starting price of $999.99 in Q1 2027
• It is the only computing product in an AI-branded launch with no TOPS figure and no Copilot+ claims as part of its published spec sheet
• The laptop does make compromises, skipping Thunderbolt and USB4 connectivity, WiFi 7 at half the channel width of the Arm-based ThinkBook 14 Gen 9 it launches alongside, and an HDMI 2.1 port that caps out at a 4K 60Hz output

Lenovo announced the ThinkBook 14x Gen 2 at Innovation World during IFA 2026: a 990g notebook running Intel Core 7 processors from the Series 3 generation, starting at $999.99 and available for purchase in Q1 2027.

It arrives as part of a series of launches focused on what Lenovo calls "Hybrid AI for Business," but it's the only computing product in the press release that makes no AI claims out of the box.

It offers excellent general-purpose compute capabilities but cuts back in some key connectivity areas that might deter power users.

A performant, lightweight option that cuts back on some key features

The higher-end version of the Lenovo ThinkBook 14x Gen 2 offers a 14-inch 16:10 2.8K OLED running at 120Hz with 100 percent DCI-P3 coverage, 500 nits, and a bent anti-glare finish, which is an excellent offering in a market teeming with cheaper options. For those unwilling to pay the premium it will inevitably command, the ThinkBook 14x also offers a WUXGA LCD alternative with the same 120Hz refresh rate, 100 percent sRGB, 400 nits, and optional touch.

Lenovo's press release says the ThinkBook 14x Gen 2 "delivers premium mobility in an ultra-thin, lightweight design built for everyday productivity," and it is hard to argue otherwise for a laptop that holds its own in a market that is getting pricier thanks to soaring component costs.

It does, however, make significant cuts to get there; the most prominent is the lack of Thunderbolt or USB4 ports, despite sporting a CPU (Intel Core Ultra 7 Series 3) that supports them. It also cuts Wi-Fi support to 160MHz versus 320MHz on the accompanying ThinkBook 14 Gen 9, effectively halving its maximum data throughput.

Despite this drawback, it has two USB-C ports that currently offer USB 3.2 Gen 2 connectivity at 10Gbps, which should suffice for most mainstream users even if the four-Thunderbolt-lane standard of Intel's Core Ultra Series 3 platform controller remains unused.

It also comes with an 18Gbps HDMI 2.1 port (TMDS) that caps maximum output at 4K 60Hz, which is arguably a lesser concern for a laptop that lacks Thunderbolt support, which already limits its potential to be paired with an eGPU and is unlikely to offer much to gamers, who would otherwise be the primary audience for a high refresh rate at a 4K resolution.

Lenovo has not published the AI TOPS capabilities of the as-yet-unnamed Core Ultra 7 Series 3 CPU it is pairing with the ThinkBook 14x Gen 2, even as the ThinkBook 14 Gen 9, announced alongside it features a Snapdragon X2 Plus with an explicitly stated 80 TOPS AI performance by Qualcomm and Copilot+ support out of the box.

One could argue that the AI aspect being skipped is intentional, and the ThinkBook 14x Gen 2 caters to a vastly different crowd, something that is easy to reconcile with, but the compromises in connectivity hit hard for any power user looking for a capable but feature-packed 14-inch laptop in the segment even if Lenovo has indirectly made it abundantly clear about what its target audience is: casual users that need a thin, but performant 14-inch laptop.

• Lenovo announced the ThinkCentre X Ultra at IFA 2026, a 1.6-liter desktop with its max configuration offering up to AMD's top-of-the-line Ryzen AI Max+ PRO 495 APUs
• It launches in November 2026, starts at $3,699, and comes with up to 128GB of unified memory under the hood
• Users can configure four of these together to leverage as much as 512GB of unified memory for local needs

Lenovo announced the ThinkCentre X Ultra at its Innovation World event during IFA 2026: a 1.6-liter desktop built around AMD's Ryzen AI Max+ PRO 495 with up to 128GB of unified memory, starting at $3,699, shipping in November 2026.

The company calls this offering a new class of desktop for agentic AI, and it can be configured to cluster up to four systems, effectively pooling processing power and memory to run large AI models locally.

The added compute and memory capacity lets users run longer context windows and multi-agent workflows without turning to cloud-hosted alternatives, while keeping data localized.

Plenty of scalable power under the hood

The Ryzen AI Max+ PRO 495 is the flagship of AMD's Gorgon Halo refresh, pairing 16 Zen 5 cores with a 40-compute-unit Radeon 8065S integrated GPU and an NPU rated at roughly 55 TOPS. Its defining trait is the shared memory pool: the CPU and GPU address the same LPDDR5X memory under the hood, so it can effectively hold and run models that a discrete, consumer-grade GPU with 24GB or 32GB of VRAM cannot.

Lenovo pairs it with its enterprise-class ThinkShield security, AMD's PRO technologies, and DASH out-of-band manageability, plus the self-healing BIOS that runs across the ThinkCentre line, and integrates the AMD Ryzen AI Developer Center for preconfigured tools, models, and workflows across both Windows and Linux.

The Yoga laptops unveiled alongside it also reach 128GB of unified memory, but they don't offer the ability to cluster together for more memory or the additional hardware you may not need for a localized AI server. The price might be an issue for many looking to hook up four of these devices together; however, assuming a device cost of $3700 for a 128GB model, one would need nearly $15,000 before factoring in connectivity costs to build a 512GB memory cluster based on the ThinkCentre X Ultra.

Lenovo has not, however, said how connectivity will work, what bandwidth units will have when running four in tandem, or what the performance implications will be for such a configuration, even as its listed specs include dual Thunderbolt 4 ports rated at 40Gb/s each. The larger memory pool also might not translate to performance; while AMD's 16-core Zen 5-based configuration is fast as processors go, the integrated GPU is considerably slower, even when running together, versus something like an Nvidia H100 that can be rented for a few dollars an hour online.

For those looking for something visually attractive that doubles as a monitor for what the system is up to, the ThinkCentre X Ultra also has adaptive lighting under the hood that provides real-time visual feedback. However, whether it stacks up two months later, as an increasing number of Ryzen AI PRO 400 APUs make it to developers and AI users across the world, joining an ever-increasingly competitive fray, remains to be seen, even if you can rig four of them together, in addition to Lenovo's enterprise offerings being integrated in tow.