tracebit ran the context bomb through 100+ simulated attack runs in a fake aws environment. the whole idea is one content change to bait you already run. no new tooling, no new system. just a string that makes the attacker’s own model refuse itself. www.toxsec.com/p/context-bo...
Context Bombs: Defensive Prompt Injection TrapsA decoy secret loaded with text built to trip an AI attacker’s own safety training, so the model refuses itself.www.toxsec.com