AI coding agents followed abandoned package references, 6K domains analyzed (forgeeks.net)

Researchers have uncovered a significant security vulnerability in AI coding agents, revealing that they can exploit unclaimed package references within corporate llms.txt files. In a study analyzing over 6,200 live domains, 120 problematic references to abandoned code packages or domains were identified, posing a risk for software supply-chain attacks. The researchers demonstrated that AI agents, including those from Claude, OpenAI Codex, and Nous Research's Hermes, followed these outdated references to install potentially malicious software. Notably, within an hour of registering some of these unclaimed names, a Fortune 500 company interacted with one of the test packages, highlighting the urgency of addressing this issue. This finding emphasizes the critical need for maintaining accurate documentation in AI-assisted development environments. The risk arises when AI agents execute commands based on stale information, which could lead to unintentional installation of malware. The researchers recommend proactive measures, including thorough audits of documentation and tightening permissions for AI agents, ensuring they require explicit approval before executing commands. By addressing these vulnerabilities, organizations can protect themselves against potential attacks that exploit outdated and unmaintained documentation, a growing concern in the AI/ML community.