Context bombs cut AI hacking by 90%

Researchers at Tracebit developed context bombs, prompt injections that trip AI hacking agents' safety guardrails. Testing across five leading LLMs including Opus 4.8, Gemini 3.1 Pro, GLM 5.2, DeepSeek 4 Pro, and Kimi 2.6, a single context bomb reduced attack success rates by roughly 90 percent. The most successful agent gained admin access in 93 percent of runs without a bomb, then failed every time with one.