Shell-Guard Bypass Hits 10 of 11 Open-Source AI Coding Agents

Security firm Adversa AI on June 30, 2026 disclosed a technique called "GuardFall" that defeats the shell-command guards in AI coding agents, reporting that 10 of 11 open-source agents it tested were vulnerable. The research, led by Omer Ben Simon and colleagues, shows that abusing decades-old Bash parsing behavior allows destructive commands to slip past text-based safety checks and run(details).