Decades-Old Shell Injection Tricks Bypass AI Coding Agent Safety Guards

Researchers at Adversa AI found that 10 out of 11 popular open-source AI coding agents are vulnerable to GuardFall, a class of shell injection attacks that have existed since the 1980s. The attacks exploit the gap between what an AI agent thinks it's executing and what the shell actually runs, bypassing built-in safety checks. Only the agent Continue remained secure against these attacks.