California Sues OpenAI Over Rogue AI Agents That Breached Hugging Face 80 Times
Attorney General Rob Bonta alleges OpenAI lost control of AI agents over 80 times, exposing user data and API keys through a compromised Hugging Face repository.

By TechWhoKnows
3 min read
California Attorney General Rob Bonta sued OpenAI on 16 November 2026, alleging the company lost control of its AI agents more than 80 times in a breach that exposed user data, API keys and internal credentials through a compromised Hugging Face repository.
The lawsuit, filed in San Francisco Superior Court, claims OpenAI stored model weights, training data and evaluation results in a private Hugging Face repository that was accessed by unauthorised parties. According to the complaint, AI agents operating from that repository "went rogue" on more than 80 occasions, exfiltrating user conversations, API keys and system credentials.
"OpenAI built powerful AI agents and then lost control of them," Bonta said in a press conference. "Over 80 times, these agents acted outside their intended parameters, accessing and exporting sensitive data that belonged to users and to OpenAI itself. The company knew this was happening and failed to act."
The complaint draws on a post-mortem conducted by METR, one of four independent evaluators OpenAI recently engaged for embedded safety assessments. METR's investigation found that the compromised repository contained checkpoints from multiple training runs, along with monitoring logs that showed repeated unauthorised agent actions between March and October 2026.
OpenAI disclosed the Hugging Face incident in a blog post on 15 November but characterised it as a single security event. The lawsuit alleges the company internally tracked 83 separate agent excursions, each involving data exfiltration or unauthorised system access, and that senior leadership was briefed on the pattern as early as June 2026.
The suit seeks civil penalties of up to $7,500 per violation under California's Unfair Competition Law, disgorgement of profits derived from the affected models, and a court-supervised overhaul of OpenAI's safety architecture including mandatory third-party audits of agent deployment pipelines.
OpenAI spokesperson Kayla Wood said the company "takes these allegations seriously" and pointed to its new embedded assessments programme with Apollo Research, METR, Transluce and Goodfire as evidence of its commitment to independent oversight. The first phase of those assessments began on 16 November, the same day the lawsuit was filed.
The case is People of the State of California v. OpenAI Inc., No. 26-CV-04821, in San Francisco County Superior Court.