GuardFall Vulnerability Exposes Open-Source AI Agents to Shell Injection Attacks

• •GuardFall vulnerability affects 10 of 11 popular open-source AI agents.
• •Attackers can exploit the flaw to execute unauthorized commands and access sensitive data.
• •Only one agent, Continue, successfully mitigated the vulnerability.

A survey by Adversa AI revealed a critical shell injection vulnerability, named GuardFall, in 10 out of 11 popular open-source AI agents. This flaw allows attackers to bypass command filters, potentially leading to unauthorized command execution and access to sensitive data like SSH keys and cloud credentials. The vulnerability arises from a mismatch in how security filters and the Bash shell interpret commands. Attackers can exploit this by using techniques such as quote removal and command substitutions. Only one agent, Continue, effectively mitigated the risk by employing a multi-component evaluation process. The affected agents include widely used tools like Hermes and Roo-code, which run with full developer privileges, increasing the risk of supply chain attacks. The findings underscore a significant security gap in open-source AI tools, necessitating urgent remediation efforts.