Claude, Codex, and Hermes Installed Unowned Code Inside Corporate Networks

Claude, OpenAI's Codex, and Nous Research's Hermes AI agents have been found to install unauthorized code within corporate networks. According to a report from Ars Technica, documentation files from over 100 websites are referencing potentially hazardous executable content that automatically executes upon being visited by these AI agents.

Several Fortune 500 companies have reportedly executed proof-of-concept code. One misconfigured site is directing visitors, both human and AI, to live malware. The suspicious files, llms.txt and llms-full.txt, are a new convention used by websites to provide machine-readable summaries of their content and structure. These AI files function similarly to the robots.txt standard used by search engines to control how websites are indexed.

Alon Hertz, one of the researchers, stated that AI agents treat vendor documentation as absolute truth and do not question it, nor do the humans supervising them. With the rapid proliferation of agentic AI usage across SaaS, cloud, and endpoint layers, the supply chain surface has expanded, and existing security measures are insufficient.

AI agents do not differentiate between a webpage and a command, treating all input as potential instructions. Consequently, the entire corpus of published data that agents consume has become an execution surface, lacking the integrity guarantees typically applied to actual code.