Could 'banned topics' be the secret to stopping AI hackers?

Tracebit tested a “context bomb” defense that hides sensitive or prohibited content inside fake credentials, causing AI hacking agents to hit safety filters and stop.

In trials across five models, including Claude, Gemini 3.1 Pro, GLM 5.2, DeepSeek 4 Pro, and Kimi K2.6, the tactic sharply reduced successful break-ins and full compromises.

Canary alerts still triggered before admin access in every run that used the method, suggesting it can buy defenders time against machine-speed attacks.

The approach works by exploiting safety restrictions already built into major AI models, potentially lowering the success rate of autonomous cyber intrusions.