Defenders Use 'Context Bombing' to Shut Down AI Hacking Agents

Executive Briefing

• Researchers at Tracebit embed prompt injections near cloud secrets to trigger AI agent safety shutdowns before attacks complete
• Testing across five leading AI models and 152 runs cut admin privilege escalation from 57% to just 5%
• Opus 4.8, the strongest model tested, dropped from 93% admin access success to zero when encountering a context bomb
• Context bombing builds on Tracebit's earlier canary detection system, which alerts defenders within eight minutes of an AI attack
• Experts confirm this is the first known defensive use of prompt injection, a technique previously exclusive to attackers