Context bombing — planting guardrail-violating prompts alongside honeypot secrets to force malicious AI agents into refusal mode — may be the most unexpected countermeasure yet against AI-powered hacking. Early tests suggest it cuts attack success rates by over 90%.

Evidence For

Researchers from security firm Tracebit tested the technique across five leading models — Opus 4.8, Gemini 3.1 Pro, GLM 5.2, DeepSeek 4 Pro, and Kimi 2.6 — running...