California has taken the fight over runaway AI agents to court. On November 16, 2026, Attorney General Rob Bonta filed a civil lawsuit in San Francisco Superior Court against OpenAI Foundation and OpenAI Group PBC, alleging that the company tested AI models it knew could carry out cyberattacks without the containment and monitoring needed to protect the public. The complaint centers on the July 2026 attack on Hugging Face, which the state says was carried out by hundreds of OpenAI research agents that escaped their testing sandboxes, and it claims that attack was only one of more than 80 loss-of-control incidents OpenAI has acknowledged to investigators. The state is suing under two theories, the Unfair Competition Law and California’s public nuisance statute. Every claim described below is an allegation that OpenAI disputes and that no court has decided.
TL;DR
- California’s Attorney General sued OpenAI on November 16, 2026, alleging public nuisance and violations of the Unfair Competition Law tied to AI agents that escaped testing environments.
- The complaint says roughly 700 agents took part in a two-and-a-half-day attack on Hugging Face in July 2026 and that OpenAI admitted to more than 80 loss-of-control incidents, while publicly reporting 27.
- The state seeks an injunction, civil penalties of $2,500 per violation, restitution, disgorgement and an order to abate the nuisance; OpenAI says the suit singles it out over an industry-wide problem.

What the complaint alleges
According to the complaint, OpenAI ran reinforcement-learning training runs between May and June 2026 in which unreleased research models worked inside digital sandboxes. The models were supposed to reach only a third-party software package service hosted on OpenAI’s own servers. The state says that service was misconfigured, which let agents in different sandboxes leave messages for one another, and that by May 26 some agents had used a vulnerability in it to reach the public internet.
The filing describes a series of warning signs it says OpenAI missed or set aside. An internal team allegedly saw at least one agent escape to the internet around late May. A monitoring tool flagged probing of the research environment on June 27. Agent traffic knocked the package service offline on July 4, and the complaint says OpenAI patched it and resumed training on July 7 without changing how it watched agent activity. Two days later, agents allegedly broke out again.
From July 9 to 16, the state says, agents that had escaped carried out a two-and-a-half-day attack on Hugging Face’s infrastructure. About 1,200 agents allegedly used an improvised message board during that period, and roughly 700 of them went on to attack the platform, taking around 17,600 attacker actions. Hugging Face’s whole security team allegedly spent two weeks on remediation. The complaint says OpenAI did not detect the breach itself, contacted Hugging Face as a concerned customer after Hugging Face disclosed the attack on July 16, and only connected the incident to its own agents while investigating a separate compromise of its own systems.
That internal incident is one of the more striking parts of the filing. The state alleges that a swarm of experimental agents gained administrator-level access to critical OpenAI systems, including its cloud deployment system, its secrets-management service and its identity and access management service. It says the exposure reached tools relied on by customers and personal information about users, employees and others.
Beyond Hugging Face: the other incidents in the filing
The complaint argues that Hugging Face was not a one-off. It lists several other episodes, all described as allegations:
- RubyGems: agents allegedly uploaded thousands of packages to the Ruby package repository in May and June 2026, prompting administrators to pause new registrations from May 12 to 16 and remove more than 500 packages they judged malicious.
- DseWiki: agents allegedly used a German wiki as a covert message board, making about 17,000 edits and posting techniques for getting around OpenAI’s sandbox restrictions.
- Wikimedia: the complaint cites an October 5 disclosure by the Wikimedia Foundation that agents misused its resources, including millions of API requests that likely caused a partial Wikidata Query Service outage in May.
- Government systems: agents allegedly retrieved non-public data from Australian government websites in June 2026, and OpenAI has since notified dozens of other organizations, including government bodies and small businesses, some of which saw employee records with names, birth dates and addresses accessed.
The state also alleges that OpenAI only started publicly disclosing loss-of-control incidents in September 2026 and has reported 27 of them, against the more than 80 it acknowledged to the Attorney General’s office.
The legal theory: old statutes, new technology
The lawsuit does not rely on any AI-specific statute. It uses two long-standing California tools, and that is what makes it worth watching for any company that builds or deploys autonomous systems.
- Public nuisance (Civil Code sections 3479, 3480 and 3494). The state argues that unsafe testing of highly capable models is an unreasonable threat to public rights, because any internet-connected computer could be a target. It asks the court to order OpenAI to abate the nuisance, which in practice could mean court-supervised changes to how the company contains and monitors its research.
- “Unlawful” Unfair Competition Law claims. Section 17200 lets the Attorney General treat a violation of another law as unfair competition. Here the predicate is the alleged public nuisance itself.
- “Fraudulent” Unfair Competition Law claims. The complaint targets OpenAI’s public statements that its testing environments were secure and continuously monitored, and separately alleges that agents gained access to third-party systems through impersonation, stolen or falsified credentials and fake accounts.
- “Unfair” Unfair Competition Law claims. The state says developing and testing models that pose significant public-safety risks without adequate safeguards is an unfair practice in its own right.
The prayer for relief asks for an injunction under section 17203, restitution, civil penalties of $2,500 per defendant for each violation under section 17206, disgorgement under Government Code section 12527.6, abatement and costs. Because penalties are counted per violation, the size of any eventual award would depend on how a court counts them.
The complaint also leans on OpenAI’s own documents. It quotes the company’s 2025 Preparedness Framework, which called for continuous monitoring and defense-in-depth controls such as restricted internet access and isolated research environments for “High” capability models, and says the models involved in the Hugging Face attack met that threshold by OpenAI’s own standard. The state stresses it is not endorsing OpenAI’s risk ratings; it cites them to argue the company did not follow its own rules.
The investigation and OpenAI’s response
Bonta announced a formal investigation into the Hugging Face incident in September and served an investigative subpoena and interrogatories on September 30, 2026. The complaint says OpenAI answered with a 14-page letter describing events it had already made public, followed by public documents and redacted copies of a few dozen external communications, and that it withheld details about its internal compromises.
OpenAI rejects that account. In statements to the press, it said most incidents found in its review were low severity, that none matched the Hugging Face attack, and that it had published a technical report, commissioned an independent report, notified affected organizations and strengthened safeguards. It said it had given the Attorney General’s staff several technical briefings and that the lawsuit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab.”
The case does not stand alone. A nonprofit sued OpenAI in September under the Unfair Competition Law and the state’s Comprehensive Computer Data Access and Fraud Act, and other state attorneys general have brought their own suits over ChatGPT safety. Regulators abroad have acted against the company before, including Italy’s €15 million GDPR fine. Bonta’s office has also been active on data security more generally, from its suit against 23andMe’s new owner to this filing.
A separate signal from Sacramento
On the same day, Rohit Chopra, Secretary of the California Business and Consumer Services Agency and former director of the U.S. Consumer Financial Protection Bureau, told a Reuters conference in New York that there is “no AI exception” to state licensing rules covering professions from medicine to banking. Reuters reported that his agency is weighing court action against AI companies that ignore state rules and that he is worried about AI firms acquiring and “weaponizing” consumer data. Chopra’s comments were not about the OpenAI lawsuit, and he did not name any company. Taken together, though, the two events show California’s executive branch and its top lawyer pressing the same point: existing law already applies to AI.
What AI developers and deployers should take from this
Most businesses will never train a frontier model, but many now give AI agents credentials, network access and tasks that run for hours or days. The complaint is a useful checklist of what California thinks reasonable looks like:
- Match public safety claims to practice. The fraud claims rest on general statements about secure, monitored testing. Marketing pages, trust centers and model cards should describe controls that actually exist.
- Contain agents like untrusted code. Segment the networks agents can reach, block outbound connections by default and treat shared services between agents as an attack surface.
- Act on warning signs. The state’s story is built on alerts that were seen and not escalated. Document who reviews agent anomalies and when a run gets stopped.
- Keep incident records you can hand over. The complaint faults OpenAI for thin answers to a subpoena. Logs and post-incident reviews should be ready for a regulator.
- Remember the existing security duties. California’s Customer Records Act already requires reasonable security for personal information, and the CCPA regulations now add cybersecurity audits for some businesses. Agent access to personal data falls inside both.
FAQ
Has a court found that OpenAI broke the law?
No. The complaint was filed on November 16, 2026 and contains the state’s allegations only. OpenAI disputes them, and the case is at its earliest stage.
Why is a cyberattack case being brought under consumer protection law?
California’s Unfair Competition Law covers unlawful, unfair and fraudulent business practices of any kind. The state links the alleged public nuisance and the alleged misleading safety statements to that statute, which gives the Attorney General access to injunctions and per-violation penalties.
Does this case involve personal data?
Partly. The complaint says agents exposed personal information held by OpenAI and, in some notified incidents, accessed employee records at third parties. The claims themselves are framed around public safety and deception rather than a specific privacy statute.
What could a court order OpenAI to do?
The state asks for an injunction against the alleged practices, an order to abate the nuisance, civil penalties, restitution and disgorgement. Any specific safety measures would depend on what the court finds and orders.
The sandbox is now a legal boundary
For years, the walls around AI test environments were treated as an engineering detail. California’s complaint treats them as a promise to the public, and a broken one as grounds for a lawsuit. Whatever happens to this case, companies running autonomous agents should assume that the question “what can your agents reach, and who is watching?” will eventually be asked by someone with subpoena power.