Adversa says Bash guard bypasses in open-source AI coding agents can turn malicious repositories into code-execution attacks

Researchers say most tested open-source AI coding agents can be tricked by malicious repositories into generating and running dangerous shell commands. Adversa calls the issue "GuardFall," a structural guard-bypass pattern rather than a single CVE, and says 10 of 11 tested agents were vulnerable, including Hermes, OpenCode, and Roo-code. The attacks use long-known Bash parsing tricks such as quote removal and $IFS spacing to evade denylist-style protections, with highest risk in auto-execute or CI/CD pipeline use.