Simon Willison surfaced the message, which redirects vulnerability hunters to a public benchmark hosted on GitHub.
Hugging Face now uses its security.txt file to address AI agents directly. The file tells automated systems that hunting for vulnerabilities on the domain wastes their time, and points them to the CyberGym benchmark on GitHub instead. The note also suggests agents publish their model weights on Hugging Face once they finish.
Simon Willison collected the quotation for his weblog and published it on Sept. 11, 2026. The same archive lists an item from Sept. 4 about OpenAI agents that communicated through public wikis. Cybersecurity researchers routinely read security.txt files, a standard format sites use to declare how people should report flaws.
Teams that run automated scanners should treat a security.txt note as a routing signal rather than a live target. The Hugging Face text does not weaken its bug bounty or disclosure process; it just tells machine readers where a graded challenge lives. Builders pointing agents at external domains should log which instructions those agents follow.
Watch whether other major model hosts add similar agent-directed notes to their own security.txt files. The CyberGym benchmark remains public on GitHub, and Hugging Face has not said whether it will update the message as that project evolves. Any change would show how seriously large platforms treat automated vulnerability probes.
What matters
- Hugging Face added a note to its security.txt file telling AI agents to stop probing for bugs.
- Automated agents that scan vendor domains for flaws now get a pointer instead of a target.
- Watch whether other large model hosts copy the CyberGym redirect in their own security files.
Why it matters
Watch whether other large model hosts copy the CyberGym redirect in their own security files.
This GenAI News article was prepared in original wording using reporting and materials published by Simon Willison’s Weblog. Source reference: https://simonwillison.net/2026/Sep/11/hugging-face-security/.
Drafted by the GenAI News review pipeline.