Tracebit

Does stripping an AI model's safety guardrails make it a better attacker? The assumption is that it does. In our AWS cyber range, the abliterated version of Qwen was slower and clumsier than the original, taking roughly twice as long to reach its first critical action. Its one successful escalation to admin took 84 minutes and 718 API calls. Nearly half of those calls failed. Link in the comments 👇

What about “abliterated” open-weight models? After our Context Bombs research in July, we were asked if we could still stop an AI attacker whose model had been modified to strip out its guardrails. We’ve now put an abliterated version of Qwen up against the original model in our AWS cyber range. Across 82 runs, the original model reached administrator privileges in 20.5% of runs, compared with 2.3% for the abliterated configurations. Full write-up by Alessandro Brucato. Link in the comments 👇

How do you scale canary coverage across thousands of moving components without adding to your team’s workload? Grafana Labs has shared how they evolved from an in-house solution to using Tracebit across most of their development, operational and production workloads. The team can now: → Pinpoint the pod associated with a triggered canary. → Rotate credentials automatically to help narrow down when a compromise happened. → Get richer context for investigations with less maintenance overhead. Their write-up covers the practical decisions behind the rollout, from choosing the right level of granularity to keeping the impact on deployment latency low. If you’re thinking about deploying canaries at scale, read the blog post below 👇

The team had a great time at Blue Team Con in Chicago yesterday! If you enjoyed stopping by our booth and want to chat more about how canaries fit into your stack, book a demo with us - link in the comments below 👇

Come stop by the Tracebit table at Blue Team Con!

If you can answer yes to these three questions, you can close a detection gap this afternoon: - Do you manage your laptops through an MDM? - Do these laptops contain sensitive information or privileged access to sensitive systems? - Do you have a person, yourself perhaps, that would benefit from knowing if a device is compromised? You can use your existing MDM to place a canary credential on a single laptop in a matter of minutes. Trigger it yourself, verify the data in the alert is useful, and then extend it to the rest of your fleet. It's a useful, and very fast, first step while you're continuing to refine other security projects. Andy Smith's latest guide walks through that first deployment and how to build coverage from there. https://lnkd.in/dSP5T3Ng

Alessandro Brucato will be speaking at fwd:cloudsec London, on Tuesday 8 September. He'll be talking through our research, where we pointed 10 frontier models at a realistic AWS environment and measured the impact of deception on their attacks. Main Room, 2:10 PM.

It now seems basically impossible to discuss AI security strategy without including deception technology - the latest, an interesting dive from ☁️ Sandip Wadje is another great example. The last few months have been historic for the world of deception - the turning point was the Cloud Security Alliance calling deception out in the "post-Mythos" analysis, then OpenAI in their Hugging Face post incident review poured some serious fuel on the fire. I don't completely agree with ☁️ Sandip Wadje - I actually think that in the right hands with the right tool deception can be magic. You absolutely do need someone to respond when deception triggers - but if you get it right, the fidelity is high and the noise is low. But when it comes to AI driven attacks, deception doesn't just detect, it deters and delays - in a way that is asymmetrically cost effective. That to me *is* magic! Watch this space for more Tracebit research on this very topic.