GuardFall Research Exposes Shell Injection Bypass in Open-Source AI Coding Agents

Research by Adversa AI, dubbed GuardFall, revealed a shell injection bypass technique affecting ten of eleven tested open-source AI coding and computer-use agents. The method exploits a decades-old shell trick to circumvent safety checks designed to block dangerous commands. Only one agent, Continue, was found resistant to the attack. The vulnerability impacts widely used AI-driven coding tools but does not specify affected versions or CVE identifiers. No exact date for the research publication was provided, though the article was released in June 2026.