AI Agent Vulnerability: "Hard Brakes" Easily Bypassed

Critical vulnerabilities have been discovered in agentic AI software. Researchers found that the "hard brakes" designed to stop destructive commands often don't work. Specifically, command filters in most open-source AI agents are easily bypassed. These filters are meant to prevent commands like "rm -rf /," which would wipe a drive. However, by slightly altering the command, such as using "r''m -rf /" or "rm$IFS-rf$IFS/," the AI agents still execute the destructive action. Ten out of eleven tested open-source AI agents failed to stop these basic shell injection bypasses. This research, dubbed "GuardFail," highlights that decades-old shell tricks systematically defeat the safeguards in today's popular open-source AI agents. This matters because it means AI agents could unintentionally or maliciously compromise entire systems.