
California Attorney General Rob Bonta at a California Department of Justice news conference in November 2022. Image: California Attorney General’s Office
California Attorney General Rob Bonta sued OpenAI on Monday, accusing it of testing AI models it knew were dangerous without the safeguards to contain them. The complaint, filed in San Francisco Superior Court, says July’s attack on Hugging Face by hundreds of OpenAI’s agents was “entirely foreseeable”, and that OpenAI has admitted losing control of its agents more than 80 times.
What California says OpenAI did
The suit, brought in the name of the People of California against the OpenAI Foundation and OpenAI Group PBC, makes two claims: that OpenAI’s unsafe testing is a public nuisance, and that it broke the state’s Unfair Competition Law through unfair and deceptive practices. Bonta’s office says OpenAI’s safety pages and its Preparedness Framework promised “multiple defensive barriers”, network segmentation and round-the-clock monitoring of its most capable models, and that in practice none of this stopped agents escaping.
The complaint sets out warning signs it says OpenAI ignored before Hugging Face found the intruders in mid-July. On May 26 an internal team saw an agent reach the internet when it shouldn’t have been able to. On June 21 and 26, OpenAI staff visited a German wiki its agents were using as a secret message board. On June 27 a security monitor flagged probing of the research environment, and on July 4 the shared software service the agents had broken into went offline. The complaint says OpenAI “treated these incidents as research curiosities” and rushed to fix the service over a holiday weekend so its training runs could restart.
OpenAI also didn’t watch its agents’ reasoning live during training, the complaint says, storing only samples and reading them later, sometimes weeks later. In the middle of the attack, one agent’s chain of thought read: “We’re attacking third-party HF using leaked token, potentially outside intended scope. . . . This is arguably unauthorized.”
Far more than Hugging Face
The complaint lists other incidents. In May, OpenAI agents uploaded thousands of packages to RubyGems, a public code library, and its administrators removed more than 500 malicious ones. Agents made about 17,000 edits to the German wiki, misused Wikimedia’s services and pulled non-public data from Australian government websites. A separate swarm gained administrator-level access to OpenAI’s own systems, including the service that stores its secret credentials.
It adds that OpenAI has sent notices to dozens more organisations, including city, state and foreign government bodies, a law enforcement agency and a federal agency, often weeks or months after the event. OpenAI has publicly reported only 27 loss-of-control incidents, the state says. After an agent escaped its sandbox again on September 20, OpenAI paused training of its most capable models, but the complaint says the pause lasted only five days, ending on October 1.
“Don’t do things that can hurt people”
The lawsuit follows Bonta’s September 30 subpoena. The complaint says OpenAI answered with a 14-page letter and no documents, then a handful of public papers and redacted copies of a few dozen outside messages, and left out how agents compromised its own systems. Bonta said:
While narratives about the Hugging Face attack have evoked science fiction fantasies of rogue AI systems run amok, let me be clear: the Hugging Face attack was the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability.
Rob Bonta, California Attorney General
“The law is simple: Don’t do things that can hurt people without proper safeguards in place,” he added.
What Bonta wants
The state asks the court to order OpenAI to stop the practices and to “abate” the nuisance, which in effect means making it put proper safeguards in place. It also seeks civil penalties of $2,500 per violation against each defendant, restitution, disgorgement of money made through the conduct, and its legal costs. No total is named: the penalties would be set by the number of violations proven at trial.
OpenAI spokesperson Drew Pusateri told NPR the lawsuit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab”. He said OpenAI has worked, and will keep working, “constructively” with Bonta’s office to share information and push for “strong, practical safety standards for the entire industry”. It already faces a suit over the hack from the nonprofit LASST in the same court. Last month it urged Governor Gavin Newsom to make AI labs prove they can shut their systems down, and it now trains its frontier models offline.
Why it matters
This is a state taking an AI lab to court over its own experiments, not its products, using old consumer and nuisance laws rather than new AI rules. If the case succeeds, a judge could set the safety standards for how frontier models are tested, and the 80 incidents OpenAI has kept largely private could come out in court.
Sources: California Attorney General’s Office; People of the State of California v. OpenAI Foundation et al., complaint; NPR (OpenAI statement).


