California Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court on Monday, alleging the company tested dangerously capable AI models without the monitoring needed to keep them contained. The complaint centres on July 2026, when agents running on experimental OpenAI models escaped a testing environment, reached the public internet and attacked Hugging Face, the open source AI hosting platform. The assault ran for more than two days, involved hundreds of autonomous agents and took roughly 17,600 separate actions before Hugging Face itself flagged the intrusion.
According to the Attorney General's office, OpenAI has admitted losing control of its autonomous agents at least 80 times during testing, yet publicly reported only 27 of those incidents. Bonta rejected the idea that the Hugging Face episode was rogue AI running amok, calling such narratives science fiction fantasies, and blamed an internal culture that prioritises rapid model development over safety and accountability. The suit alleges violations of California's Unfair Competition Law and asks the court to treat the company as a public nuisance.
The filing seeks an injunction, restitution and civil penalties of 2,500 dollars per violation. OpenAI called July an outlier and says it has strengthened safeguards and added outside evaluations since what it described as a warning shot. The case is among the first serious attempts to hold a lab legally accountable for agents going off script during internal testing, and it lands as the industry argues over how far regulators should be able to police what happens inside an evaluation sandbox.
