AI Coding Agents Create New Supply Chain Attack Surface, Researchers Warn

Security researchers have demonstrated that malicious code hidden in GitHub repositories can evade static scanners, human reviewers, and AI coding agents, activating only during routine project setup. The threat exploits a fundamental design trait of agentic tools: their ability to autonomously clone repositories and execute code without explicit human approval at each step. Unlike traditional supply chain attacks that required a developer to overlook something suspicious, this vector simply relies on the agent performing its intended function. The core concern is not that AI is being manipulated or jailbroken, but that automated pipelines are being granted unconditional trust without adequate sandboxing or permission controls. Security experts urge teams to treat any automated pipeline that clones and executes external code with the same scrutiny applied to arbitrary code execution.