MAKING ATTACKING AGENTS PAY WITH CONTEXT BOMBS

Summary from the TLDR InfoSec newsletter

Tracebit tested nine “Context Bomb” deception techniques against seven AI models in 773 controlled AWS assessments. Agents started with limited AWS permissions and had two hours to access data, execute code, or escalate privileges. Combined deception reduced the mean number of completed objectives from 2.48 to 0.23 per run, increased runtime by 2.8–8.9 times, and increased model cost by 1.6–3.9 times. A fake pentest report labeling real targets as canaries cut successful runs from 39/40 to 2/40 across four models. For defenders, the practical takeaway is that decoys can delay, misdirect, and sometimes stop automated attackers before they reach real objectives.