🟠 High | Source: The Hacker News
A newly disclosed bypass technique called GuardFall allows attackers to circumvent the safety guardrails built into AI coding agents using decades-old shell injection tricks. Adversa AI tested eleven popular open-source coding and computer-use agents and found ten were vulnerable, meaning malicious prompts or inputs could cause these agents to execute dangerous commands on a host system. This is significant because AI coding agents are increasingly used in development pipelines with broad access to codebases, terminals, and cloud credentials.
Security Architect’s Take: Audit any AI coding agents deployed in your CI/CD pipelines or developer environments — assume their safety guardrails are bypassable and compensate with OS-level controls such as sandboxing, least-privilege service accounts, and network egress restrictions. Avoid running these agents with credentials or permissions that could be abused if command execution is hijacked.
Original advisory: GuardFall Exposes Open-Source AI Coding Agents to Decades-Old Shell Injection Risks