Cogent

Recent reports of agents circumventing sandbox controls reinforce the guiding principle behind how we build cloud agents at Cogent: we cannot rely on agents choosing to behave. We assume any agent may be compromised and design for that possibility. That matters when agents handle sensitive customer security data and work without real-time human supervision. At Cogent, trusted infrastructure outside the agent creates isolated sandboxes, applies tenant-isolated, deny-by-default network policies, authorizes tool calls based on agent identity, and uses a credential proxy to enable authenticated API requests without exposing the underlying secrets. These boundaries also make useful autonomy possible. Our cloud agents keep context graphs of customer security environments current, identify exploitable attack paths, and drive internal development work from ticket to PR. Our new blog post explains what we’ve learned building this infrastructure, with interactive diagrams showing how the pieces fit together. Check out the link in the comments! Special thanks to my teammates who also worked on this infrastructure: Anirudh Ravula Jia Z. Larsen Weigle Yaoyang Lin Henry Yi Ajith Kemisetti

An agent can follow its prompt perfectly—and still take an action it shouldn’t. 👀 That’s why agent safety has to extend to the tools agents use and the actions they take. Policy enforcement at runtime is one of the most underrated pieces of getting agents right. We’ve been researching how to safeguard agents beyond the prompt. Read more here: https://t.co/DlvVOtGT7Q Where do you think the real boundary for agent autonomy should sit? 🔒

It's Welcome Wednesday! This week we're spotlighting Maliha Rabbi, who recently joined the Cogent team as our first Technical Account Manager! 🎉 Maliha comes to us from Armis and Devo, where she spent her career in cybersecurity helping some of the world's largest enterprises turn complex security tooling into real, measurable outcomes. She knows what it takes to earn a customer's trust and drive results, and she brings exactly that customer-first instinct to the teams she'll support at Cogent. Outside of work, you'll find Maliha carving down the ski slopes or chasing her next adventure somewhere new, always proudly repping Canada. 🎿🇨🇦 Welcome to the team, Maliha! #WelcomeWednesdays

An AI assistant becomes a security risk when it can do three things at once: 🔸 Access private data 🔸 Read content an attacker can control 🔸 Send information or take action outside the organization That combination is known as the “lethal trifecta.” Our latest research examines how TRACE, an external policy-enforcement engine, performed across five public security benchmarks. It also looks closely at what happened after an agent was blocked and what each policy repair cost. Read the blog in comments to learn more.

Most security leaders don’t need another vendor pitch. They need a better way to see what’s emerging before it ends up on everyone’s shortlist. That’s a big part of what we’re building at ATICO: bringing security leaders together with emerging technology, experienced operators, and each other in environments where the conversations can happen naturally. Our next one is coming up in Chicago. A small group of security leaders. A few emerging technology companies worth knowing. Good food, good drinks. If you're in the Chicago area and want a better view into what’s being built outside the usual evaluation cycle, come join us. Register today. Space is limited. → https://luma.com/g7e6cwwd Moonfort Security Cogent Brava Security #CybersecurityChicago #EmergingCybersecurity

This month’s Cogenteer spotlight is on Lauren England, our founding Customer Success Manager. Lauren joined Cogent to build: to work with technical teammates, solve hard customer problems, and help shape what customer success looks like at an early-stage AI security company. Watch the video to hear what drew her to Cogent and why she’s excited about the market opportunity ahead. If this sounds like the kind of team you’d want to build with, check out our open roles in the comments. #CogenteerOfTheMonth

How do you give cybersecurity agents enough autonomy to solve a problem—without giving them enough authority to create another? Closing an exposed network port might resolve a security finding, OR break a production application. Knowing the difference takes customer context, coordinated work across agents, and controls the model cannot override. Context-building agents map the environment; an investigation agent follows the evidence; a remediation agent acts within an approved scope. Their work needs to connect without giving every agent the same authority. The balance between autonomy and control follows the job. Investigation needs room to explore. Production changes need approval boundaries, deterministic validation, and verification that the fix worked. Those requirements shaped the Agent Development Platform we’re building internally at Cogent. Shared context, explicit versions, governed execution, and traceable results give product and engineering teams a foundation for building agents that work together without rebuilding the infrastructure for every job. The goal is an agent workforce that can take on more security work while keeping each agent’s access and authority clear—and its actions accountable. Our new post walks through the architecture behind that workforce. Link in comments. Anirudh Ravula Ajith Kemisetti Henry Yi Larsen Weigle Sagar Maheshwari Yaoyang Lin

Pentest, bug bounty, and threat-model findings often sit outside the vulnerability program, forcing security teams to reconcile reports, scanner data, remediation work, and closure by hand. Cogent Security Assessments maps those findings to affected assets and related scanner detections, bringing the evidence into one workflow. Read the product deep dive in the comments to see how Cogent carries that context from import through verified closure. Geng Sng Haris Sohail Sourabh Katti

One month later, we’re looking back at Black Hat and the people who made our booth one of the best places to be. Thank you to everyone who stopped by Zero Chill to meet the team, talk security, and cool off between sessions. See where to find Cogent next: https://lnkd.in/g4dyWHnx #BlackHat2026 #AutonomousCyberDefense

Pentest, bug bounty, and threat-model findings often sit outside the vulnerability program, forcing security teams to reconcile reports, scanner data, remediation work, and closure by hand. Cogent Security Assessments maps those findings to affected assets and related scanner detections, bringing the evidence into one workflow. Read the product deep dive in the comments to see how Cogent carries that context from import through verified closure. Geng Sng Haris Sohail Sourabh Katti