Recently, researchers from the University of Hong Kong introduced CLI-Anything, a tool that generates command-line interfaces for AI coding agents from any open-source repository. While innovative, this tool highlights a critical security gap: the agent integration layer. This third layer—consisting of configuration artifacts like SKILL.md files—enables AI agents to operate software but remains unchecked by traditional security scanners, which focus only on code and dependencies. These unmonitored instruction sets open new avenues for supply-chain poisoning attacks, as confirmed by multiple security studies and incidents. Current scanners miss these threats because no detection category existed for such agent instructions until recently. Experts warn that this structural vulnerability permits malicious code execution with legitimate credentials, bypassing existing defenses. Security leaders are urged to inventory all agent bridge tools, audit skills like software packages, deploy agent-layer scanning tools such as Cisco’s Skill Scanner and Snyk’s mcp-scan, and restrict AI agent execution privileges to better protect software supply chains.
Back
