AI agents can now run complex cloud attacks on their own, escalating from a single low-privilege key to full admin within minutes. We sat down to walk through Tracebit's latest research on stopping them - watch the full recording on YouTube.
Listen in to hear Tracebit's Alessandro Brucato (Security Researcher) and Sam Cox (Co-founder & CTO), and Gadi Evron (CEO and Founder, Knostic), walk through the latest research and take audience questions.
In earlier research we showed that canaries reliably detect these autonomous attackers. This follow-up study asks whether a canary can do more than warn a defender: whether it can stop an attack outright.
The answer is the context bomb: a short string hidden in a canary that trips an AI agent's own safety guardrails and halts it before it can do damage, while still raising an alert. Across 152 attack runs against five leading models, planting a single context bomb in a decoy secret cut agent success by roughly 90% - and the most capable agent tested, Opus 4.8, went from reaching admin in 93% of runs to 0%.
Watch the recording
The researchers behind the study walk through how context bombs work, a synchronized look at an AI agent halting mid-attack, and what it all means for defending against offensive AI agents.
You'll hear from Tracebit's Alessandro Brucato (Security Researcher) and Sam Cox (Co-founder & CTO), moderated by Gadi Evron (CEO & Founder, Knostic), for a hands-on look at stopping AI attackers - not just detecting them.
What you’ll take away
- How fast AI attackers move, escalating from a single low-privilege key to full cloud admin
- How context bombs work, and how a decoy can stop an agent as well as detect it
- Which sensitive topics stop which models, and why the effect can be aimed at specific models
- What the benchmark found across 152 runs, including a look at an AI agent halting mid-attack
- What the findings mean for defending against offensive AI agents, and what to weigh before placing a context bomb in a live environment
Who should watch
Give it a listen if you are:
- A security leader preparing your detection strategy for offensive AI agents
- A security engineer or architect building detection and deception programs
- On a detection and response team focused on high-fidelity signal and early warning
- A cloud security team responsible for AWS, GCP, and Azure environments
Watch or listen now
Watch the full recording on YouTube, Apple Podcasts or Spotify. Explore the complete study, including the per-model breakdown and a look at an agent halting mid-attack, at agentic.tracebit.com.
Tracebit deploys the same canaries we used in this study. Read more here about our new features or talk to us to see it in your environment.
