Researchers from Tracebit have discovered that placing prompt injections alongside sensitive data in cloud environments can effectively stop AI hacking agents.

By embedding forbidden commands into decoy secrets stored on AWS, defenders trigger a shutdown mechanism in large language models (LLMs), preventing them from executing harmful actions.

This technique, called 'context bombing,' forces LLMs to refuse subsequent commands, thwarting attacks that previously relied on prompt injections to exfiltrate data or gain admin access.

Testing showed that context bombing reduced successful attack rates across five leading models from 57% to 5%, with the most capable agent failing entirely when confronted with these prompts.The method builds on earlier work using 'canary' resources to detect AI threats, offering a proactive defense against agentic adversaries.This marks the first known case where defenders repurpose prompt injection vulnerabilities as a countermeasure.