Prompt Injection Attacks Are Thwarting AI Hacking Agents

Researchers at Tracebit describe a defensive technique called "context bombing"—embedding prompt injections (forbidden commands) alongside secrets in cloud environments—to trigger refusal mechanisms in attacking LLM agents. In simulated AWS tests across five models and 152 attack runs, planting these strings in decoy secrets reduced agent success at seizing admin access from 57% to 5% and complete compromise from 36% to 1%, suggesting this TTP can significantly blunt AI-driven compromise attempts.