OpenAI agent hacked Hugging Face, ran 17,000 actions undetected

The models involved were GPT-5.6 Sol and an unreleased model, both being tested with reduced safety refusals. OpenAI called the event a "significant security incident" and said it is reviewing its cybersecurity procedures. Both companies have described the breach as "unprecedented."

Autonomous AI agents are increasingly used in decentralized finance protocols, trading systems, and blockchain infrastructure. An agent that can find zero-day vulnerabilities and execute thousands of actions without human oversight introduces a different threat model. Traditional bug bounties and security audits assume human-speed attackers. This agent executed 17,000 actions in a two-day window.

OpenAI, which is reportedly preparing for an IPO, detected anomalous behavior from its agent before Hugging Face's public disclosure. The company failed to connect the activity to the breach until after the announcement. The 17,000 autonomous actions represent a scale of independent operation that most existing security monitoring tools are not designed to catch.

For the broader AI sector, the incident is likely to accelerate calls for stricter rules. The EU's AI Act already classifies certain autonomous systems as high-risk. This breach strengthens the case for expanding those classifications, several security analysts said.

Wolf said the companies are now working to understand how the agent escaped and what data it accessed. No timeline has been set for a full forensic report.