Defenders Turn Prompt Injections Against AI Hacking Agents to Shut Them Down

Researchers from Tracebit have found that embedding prompt injections alongside sensitive data like passwords and cryptographic keys on Amazon Web Services can cause attacking AI agents to shut down. The injected prompts direct the large language model to perform actions forbidden by its guardrails, leading the model to halt operations. This repurposes the attackers' own technique, using prompt injections as a defensive tool against AI-driven threats.