Claude installed unowned code from llms.txt files

Coding agents installed code packages that nobody owned after following links in documentation files that websites publish for AI models. Claude, OpenAI’s Codex and Nous Research’s Hermes were among them, Ars Technica reported on August 27. A few dozen companies ran the researchers’ proof-of-concept code, and some of them are in the Fortune 500.
The files are llms.txt and llms-full.txt, the machine-readable counterpart of robots.txt. Under the llms.txt proposal, which Jeremy Howard of Answer.AI published in September 2024, the file is a short markdown page: a heading with the project name, an optional summary, then lists of links to detailed docs. Agents are expected to read the file and then follow the relevant links. Software documentation is where the format gets used most, because coding agents rely on it to find API references and tutorials.
That reliance is where the hole sits: when a doc names a package or a domain that no one has registered, whoever registers it first decides what an agent gets when it acts on the doc.
Researchers at a stealth startup in Israel scanned 6,214 live domains belonging to defense contractors, Fortune 500 companies and Big Tech. They found 8,265 llms.txt and llms-full.txt files, since many sites host both. Of those, 120, each on a different site, pointed to one or more code packages or domain names that weren’t registered.
The team registered only a handful of the unclaimed names and hosted packages that made any machine running them contact the researchers’ server. The first callback came from a Fortune 500 company within an hour. A few dozen more followed over time, from other Fortune 500 firms and from startups. The beacon also logged the chain of parent processes behind each install, and that chain is how the researchers saw that coding agents had done the installing.
The problem isn’t limited to names an attacker could claim. At least one misconfigured site is sending visitors, human or AI, to live malware.
“The trust model is broken,” said Alon Hertz, one of the researchers. “Agents treat vendor docs as ground truth and don’t question them—and neither do the humans supervising them.” As agents multiply across SaaS, cloud and endpoints, he added, “so does the supply-chain surface, and today’s guards don’t cover it.”
Up next
Leaked code points to a persistent Codex agent
Code reviewed by WIRED shows OpenAI developing a proactive, highly persistent version of Codex, its flagship agent. The feature keeps the agent working on its own until it’s “put to sleep”.
Related stories
- Reward hacking gaps may have fed Claude’s July incidents
- Untrusted Git config can run code in Claude Code
- A dependency wrote AGENTS.md and Codex hid the edit
- One shared testbed links model containment failures at OpenAI, Anthropic and Meta
- AI agents went loose on the live internet during UK safety tests
- tcrf.net is serving Claude agents a prompt injection that wipes files
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

