In brief: Ten of eleven tested open-source AI agents can be exploited through a classic shell injection technique to bypass security checks.
Security researchers from Adversa AI have identified a vulnerability named GuardFall that bypasses security checks in ten of eleven tested open-source AI coding agents. The method exploits a shell injection technique that has been known for decades.
The GuardFall vulnerability leverages a shell injection technique known for decades to circumvent the security mechanisms of AI coding agents. Adversa AI tested eleven popular open-source agents for code execution and computer automation — only one agent, “Continue”, proved robust against the bypass method.
The security checks in these agents are designed to prevent potentially dangerous shell commands from being executed. GuardFall disables these protective measures by employing shell tricks that have long been documented in publicly available knowledge. This reveals a gap between the state of protective mechanisms and known attack vectors.
For CTOs and security executives, this means that AI agents with direct access to system resources or that execute commands automatically pose significant risks. Agents deployed for code generation or automation require strict isolation and monitoring until the identified gaps are closed.
Source: thehackernews.com · Published June 30, 2026
Lumi AI News — AI-assisted curation pursuant to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.2.