Researchers at a secretive Israeli startup scanned more than 6,214 active domains belonging to defense contractors, Fortune 500 companies, and large tech firms. Among the 8,265 llms.txt and llms-full.txt files they found, 120 pointed to unregistered code packages or domains.
To test the effect of these files on artificial intelligences, the researchers temporarily registered the unregistered names and hosted malicious code packages. In less than an hour, a Fortune 500 company contacted their server, and over time they received several more calls from other companies.
Subsequent analysis showed that AI coding agents including Claude (Anthropic), Codex (OpenAI), and Hermes (Nous Research) received and executed the unknown code. The companies had not responded to media inquiries by the time the news was released.