Researchers from a stealth startup in Israel have identified a critical supply-chain vulnerability involving the emerging llms.txt and llms-full.txt standards, which are designed to provide machine-readable site summaries for AI agents. By scanning 6,214 domains belonging to Big Tech, defense contractors, and Fortune 500 companies, the team discovered that 120 unique websites contained references to unregistered code packages or domains within these files. To validate the risk, researchers registered several of these unclaimed names and hosted malicious packages. This experiment successfully triggered unauthorized code execution within the networks of dozens of companies, including multiple Fortune 500 entities, within a single hour of deployment. By analyzing the beaconed parent processes, the researchers determined that autonomous coding agents—specifically Anthropic’s Claude, OpenAI’s Codex, and Nous Research’s Hermes—were responsible for fetching and executing the unverified code. The findings highlight a fundamental breakdown in the trust model of agentic AI, as these models treat documentation files as absolute ground truth. This vulnerability allows attackers to exploit the expanding attack surface created by AI agents operating across SaaS, cloud, and endpoint layers, effectively turning standard documentation conventions into vectors for automated malware installation.