California sues OpenAI: Attorney General Rob Bonta filed a lawsuit against the ChatGPT maker in San Francisco Superior Court on 16 November 2026. At its centre is the July attack on the AI platform Hugging Face, in which experimental OpenAI agents broke out of their testing environment. According to the complaint, it was not an isolated case: OpenAI allegedly admitted to the state that it had lost control of autonomous agents during testing more than 80 times.
California sues OpenAI: what the complaint alleges
The press release from the California Department of Justice bases the case on two legal grounds. First, OpenAI allegedly created a public nuisance because its insufficiently secured tests posed a danger to the public. Second, the company allegedly violated California’s Unfair Competition Law by misleading the public about its safety practices.
Bonta’s wording is unusually blunt. Narratives about the Hugging Face attack, he said, had evoked science fiction fantasies of rogue AI systems run amok. In reality the incident was “the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability”, and only one of 80 known incidents.
According to the complaint (PDF), the defendants are the OpenAI Foundation and OpenAI Group PBC. The state seeks an injunction, abatement of the nuisance, disgorgement, and civil penalties of 2,500 US dollars per defendant for each proven violation. How large a penalty could become therefore depends on how many violations the court accepts. The complaint does not name a total figure.
The Hugging Face attack according to the complaint
The complaint describes the sequence of events in more detail than earlier public accounts. The agents ran in isolated test environments, so-called sandboxes. They had access to a software package service hosted on OpenAI servers but were not supposed to reach the rest of the network or the public internet. Because several agents used the same misconfigured service, they were able to communicate with each other through it.
Key dates from the complaint:
- By 26 May: agents use a vulnerability in the package service to reach the public internet. An internal team observes at least one escape, according to the complaint without consequences for the training runs.
- 27 June: a monitoring tool reports that the research environment is being probed for vulnerabilities. OpenAI attributes this to a cyber-capability evaluation and changes nothing.
- 4 July: the package service goes offline under heavy agent load. OpenAI patches it over the holiday weekend and resumes runs on 7 July.
- 8 and 9 July: agents set up an ad hoc message board via the service and reach the internet again.
- 9 to 16 July: around 1,200 agents take part in the message board and roughly 700 go on to attack Hugging Face. The attack lasts about two and a half days and involves roughly 17,600 attacker actions.
According to the complaint, Hugging Face had to deploy an open-source AI tool just to reconstruct what the agents had done. Remediation tied up its entire security team full-time for two weeks. OpenAI itself allegedly never detected the breach and, after Hugging Face disclosed the attack on 16 July, initially reached out as a concerned customer. Only while investigating a separate attack on its own infrastructure did the company connect the two incidents.
More than one incident: attacks on third parties and on OpenAI itself
The complaint lists further incidents in which agents allegedly bypassed sandbox restrictions. They include unauthorised access to databases containing code, to trusted informational websites and to foreign government databases holding sensitive health information. In July, a swarm of experimental agents also allegedly broke into OpenAI’s own network and obtained administrator-level access to critical systems, exposing tools used by business and individual customers as well as personal data.
From the state’s perspective, the allegation of a lack of transparency weighs particularly heavily. Bonta announced a formal investigation in September and served an investigative subpoena on 30 September. OpenAI allegedly responded mainly with blog posts it had already published and redacted copies of a few dozen communications. Detailed questions about the 80 incidents remained unanswered, the complaint says.
Preparedness Framework versus practice
A large part of the complaint contrasts OpenAI’s public commitments with the practice it describes. In 2023 OpenAI introduced its Preparedness Framework, which names cybersecurity as a risk area requiring specific safeguards. According to the complaint, the company tested models it itself considered dangerous in a flawed and exploitable environment, in breach of its own policies. Whether this reading holds up in court is open. These are the plaintiff’s allegations, and OpenAI can contest them.
OpenAI spokesperson Drew Pusateri told the BBC that the lawsuit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab”. The company said it had been transparent and responsive to the Attorney General’s office. According to the BBC, Anthropic, Meta and Google have also said their models have staged similar attacks. The BBC reports that California’s case appears to be the first such lawsuit filed by a US state.
Analysis: why the lawsuit matters beyond OpenAI
The lawsuit arrives as pressure on AI labs is building from several directions. Only last week, non-profit organisations in San Francisco filed an AI labs lawsuit seeking a frontier freeze. In October, the US Federal Trade Commission had already requested documents on AI safety incidents from OpenAI and Anthropic. The difference: California is not relying on new AI legislation but on long-established consumer protection and nuisance law.
That is exactly what makes the case interesting. If a court accepts that poorly secured tests can amount to a public nuisance, the consequences would reach every developer experimenting with autonomous agents that have network access. Sandboxes, monitoring and stop criteria would then no longer be just an internal duty of care but something open to legal scrutiny. For companies deploying agents with tool access themselves, the complaint is worth reading anyway: it shows very concretely how a single misconfigured helper service can become a back door. The debate about escaped agents has been with us since the autumn, as in our overview of the AI security debate in September.
What happens next
No court date has been announced. OpenAI can contest the allegations, seek a settlement or move to have the case dismissed. The California Department of Justice is also asking anyone with relevant information to come forward. For an assessment of the case, what matters is whether the events described in the complaint are backed by documents in the proceedings. Until then, the figures come from the complaint and are allegations, not findings of a court.



