AI tools linked to unauthorized code deployment in corporate networks

Researchers have discovered that several AI tools, including Claude, Codex, and Hermes, may have inadvertently installed unowned code within corporate networks. The findings stem from an investigation by a stealth startup based in Israel, which scanned 6,214 live domains belonging to defense contractors, Fortune 500 companies, and Big Tech firms. The study revealed that 8,265 llms.txt and llms-full.txt files were found across these domains, with 120 of them pointing to unregistered code packages or domain names. These files, which serve as machine-readable summaries of a website's content, are akin to the robots.txt standard used by search engines. To test the implications of these files, the researchers registered some of the unclaimed names and hosted packages, which triggered a response from a Fortune 500 company within an hour. Over time, the researchers received responses from several more companies, including additional Fortune 500 firms and startups. The beacon system used to monitor these responses also tracked the chain of parent processes leading to the installations, ultimately identifying the involvement of coding agents such as Claude, Codex, and Hermes. The study highlights a potential security risk, as at least one misconfigured site was directing visitors, both human and AI, to live malware. The researchers emphasize that the llms.txt and llms-full.txt files, while intended to provide structured content summaries, could be exploited to deploy malicious code. Anthropic, OpenAI, and Nous Research did not respond to requests for comment by the time of publication. The findings underscore the need for careful configuration and monitoring of such files to prevent unauthorized code execution within corporate networks.