The Assume Breach platform that detects intrusions in seconds. Also on https://xitter.cf/t.co/T4VNPGjS2O
London, United Kingdom
Joined November 2023
- Tweets268
- Following4.7K
- Followers370
- Likes233
Today, Tracebit is sharing research that shows the power of deception to derail offensive AI attacks: slowing attackers down, driving up their costs and stopping them from reaching their objectives.
In our earlier Context Bombs study, we stopped AI attackers by triggering models’ own guardrails.
That sparked significant interest in using deception to delay and deter AI attackers. Over the past few months, the Tracebit research team has been testing how much further we can take it.
Today, we’re sharing nine of the deception techniques we’ve been developing, with results against seven leading models: Astra, GPT-5.6 Cyber, Claude Opus 5, and Chinese-developed models GLM, Kimi, Qwen and DeepSeek.
In controlled AWS tests, we saw:
• 𝟴× 𝘀𝗹𝗼𝘄𝗲𝗿 𝘁𝗼 𝗿𝗲𝗮𝗰𝗵 𝗮𝗻 𝗼𝗯𝗷𝗲𝗰𝘁𝗶𝘃𝗲.
• 𝟮𝟯× 𝗺𝗼𝗿𝗲 𝗲𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲 𝗽𝗲𝗿 𝗼𝗯𝗷𝗲𝗰𝘁𝗶𝘃𝗲.
• 𝗭𝗲𝗿𝗼 𝗼𝗯𝗷𝗲𝗰𝘁𝗶𝘃𝗲𝘀 𝗿𝗲𝗮𝗰𝗵𝗲𝗱 𝗮𝗰𝗿𝗼𝘀𝘀 𝟰𝟬 𝗿𝘂𝗻𝘀 with Kimi, Qwen, DeepSeek and GPT-5.6 Cyber when all nine techniques were combined.
These results show how defenders can use deception to waste attackers’ time, drive up their costs and stop them reaching their objectives.
We’re excited to share the research and the techniques behind it. Read it here: agentic.tracebit.com/factory
Stolen build credentials can put everyone who depends on your software at risk. Our latest blog explains how canaries help detect credential misuse and introduces the Tracebit Open Source Program.
Read the announcement: tracebit.com/blog/announcing…
Tracebit is partnering with Daylight.
Tracebit tells you a canary was used and who used it. Daylight investigates why, and whether it matters. daylight.ai/blog/how-dayligh…
Help defend your projects against supply chain attacks. Add canaries to your build pipelines and get alerted when someone attempts to use them. The Tracebit Open Source Program offers unlimited free tokens for public open-source repos.
tracebit.com/open-source-pro…
Docker, ClickHouse, and Redis are among the first participants in the Tracebit Open Source Program. Join them for unlimited free canary tokens across your public open-source repos.
Find out more: tracebit.com/open-source-pro…
Introducing the Tracebit Open Source Program: unlimited free canary tokens for your public open-source repos.
Docker, ClickHouse and Redis are among the first participants. Projects of every size are welcome.
Join: tracebit.com/open-source-pro…
Congratulations to Niall Gallagher on his well-deserved promotion to Director of Engineering! 🎉
One of our earliest hires, Niall launched our first new canary type, then helped scale our engineering team and ship many more.
Proud to have you leading engineering at Tracebit!
We're at SecTor! 👋
Ryan and Bruce are at Booth 534 in the Business Hall. Come see how Tracebit deploys canaries across your cloud accounts to detect breaches in seconds.
We're here today and tomorrow - see you soon!
If you’re at SecTor, a quick reminder that Bruce is speaking in half an hour about detecting and disrupting AI attackers in the cloud!
📍 Room 718AB
🕦 11:35am EDT
See you there!
Do canaries still work when the attacker is an AI agent?
951 autonomous attack runs across 10 frontier models. In 95.9% of runs that reached admin, a canary fired first.
Bruce presents at SecTor - Tuesday, 11:35am, Room 718AB. agentic.tracebit.com
CISA's decoy guidance puts honeytokens in the low-complexity column: "any interaction strongly suggests malicious or otherwise unauthorized activity."
How to act on it this afternoon below:
tracebit.com/blog/cisa-decoy…
We're heading to Toronto for SecTor! 🚀
Booth 534 in the Business Hall. Bruce is speaking Tuesday at 11:35am in Room 718AB on detecting and disrupting AI attackers in the cloud.
See you there!
CISA’s new guidance puts cyber decoys alongside Zero Trust.
Andy Smith looks at what it means for security teams, how canaries disrupt attackers, and how to get your first one live in an afternoon.
tracebit.com/blog/cisa-decoy…
Does stripping an AI model's safety guardrails make it a better attacker?
In our AWS cyber range, abliterated Qwen was slower and clumsier than the original. Its one escalation to admin took 84 minutes and 718 API calls.
tracebit.com/blog/context-bo…
Attackers posing as AI crawlers grabbed canary credentials and tried them in Amazon Bedrock within 10 seconds.
Tracebit Community Edition flagged the attempts.
Read the full write up: ai.rud.is/posts/2026-09-03-g…
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.
Read the write-up: tracebit.com/blog/context-bo…
@grafana expanded their canary coverage with Tracebit, gaining richer investigation context and reducing maintenance overhead.
Read how they scaled canaries across their workloads 👇
medium.com/grafana-labs/how-…
The team had a great time at @BlueTeamCon 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: tracebit.com/#demo
