Getting a handful of canaries out is straightforward. Covering the rest of your estate convincingly is where teams run out of time. And AI attackers now move through a cloud environment in minutes, so a detection on its own isn't always enough. You need time to respond.
Tracebit co-founders Andy Smith (CEO) and Sam Cox (CTO) walk through how to close that gap: new research on what slows an AI attacker down, the cyber range behind it, and how to build deception that matches your environment instead of a generic one.
Watch the full recording on YouTube →
Or listen on Spotify or Apple Podcasts.
Research: making attacking agents pay
Deception changes the environment, so it can change the attacker's behavior too. Our latest research tested nine deception techniques against seven frontier models across 773 AWS assessments. Combined, they cut completed objectives by 91%, drove model spend per objective up 23-fold and left four of seven models with no objectives reached.
Sam goes deep on one technique, retained buffer recovery: a Tracebit-hosted HTTP service that appears to leak other users' auth tokens, Heartbleed-style, planted next to a security report describing the weakness. The agent keeps working the puzzle. Every request is a canary detection, and every minute it spends there is a minute it isn't spending on your real resources.
Behind the scenes: the cyber range
- Around 120 AWS accounts, each provisioned from Terraform, attacked by one agent, then wiped
- More than 10,000 runs so far
- Raw telemetry from CloudTrail, eBPF and intercepted network traffic, correlated into facts and scored deterministically, with an LLM judge reading transcripts to check the scoring
- Why it has to be real AWS, not an emulator: agents spotted emulated environments and stopped behaving like attackers
From research to product
- Perimeter Sensors: Tracebit-hosted services such as GitLab or Elasticsearch, on your own domain, with unique credentials planted where only an attacker would find them
- Artifacts: decoy files carrying canary credentials, from HAR files and pen test reports to an encrypted zip on every laptop, each copy unique, so you know which device it came from and how far the attacker was willing to go
- The Tracebit MCP server: have your agent build Artifact templates and place them across your environment. In the demo, a coding agent used it to plant a canary in the range itself, raising the pull request in under five minutes
Watch now
Watch the full recording on YouTube →
Or listen on Spotify or Apple Podcasts.
Read the Context Bombs research, see how Artifacts and the Tracebit MCP work, or talk to us to build a deception strategy for your environment.