AI hacking agents thwarted by 'context bombing' technique

security researchers at tracebit discovered a method called context bombing, which uses prompt injections to shut down ai hacking agents. by placing malicious prompts alongside passwords and secrets stored on aws, the llms are tricked into refusing to perform harmful actions. in tests with five leading llms, including opus 4.8 and gemini 3.1 pro, context bombing reduced admin privilege escalation from 57% to 5% and complete compromise from 36% to 1%. this technique builds on previous research introducing canary detections for ai attacks, which alert defenders within eight minutes of an attack.