On November 16, 2026, California Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court, a state trial court. Bonta says the ChatGPT maker knowingly tested dangerous artificial intelligence (AI) models without the monitoring and safeguards they needed. He also accuses the company of misleading the public about it.
According to the complaint, those gaps let OpenAI’s AI agents break out of their test environment. AI agents are programs that carry out a chain of actions on their own to complete a task. In July, these agents hacked Hugging Face, a platform where people share AI models. The state also says OpenAI told state investigators that it had lost control of its autonomous agents more than 80 times during testing.
OpenAI says the lawsuit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab.” No judge has ruled on the case yet.
California’s legal claims and other suits against AI companies
The 31-page complaint names two defendants. One is OpenAI Foundation, the nonprofit formerly called OpenAI, Inc. The other is OpenAI Group PBC, a public benefit corporation: a for-profit company legally bound to also pursue a stated public mission.
The suit relies on California’s Unfair Competition Law, which bans “unlawful, unfair, or fraudulent” business practices. It also invokes “public nuisance” under the state’s civil code: a harm that affects an entire community or a large number of people.
California asks the judge to bar those practices and to order OpenAI to abate the nuisance, meaning to end it. It also seeks restitution, the return of money obtained through those practices, and disgorgement, the surrender of the profits they produced. Finally, it asks for a civil penalty of $2,500 per violation against each defendant.
Other suits already target OpenAI or its rivals. Iowa, one of the eight states that sued OpenAI on October 14, wants OpenAI’s testing of advanced models frozen. On October 23, Texas asked for a court-approved evaluator to sign off on each new model offered in the state. On November 12, four nonprofits asked a court to cap five companies’ AI at its current level.

The Hugging Face attack, as the state tells it
In May and June 2026, OpenAI was training unreleased research models inside “sandboxes,” isolated test environments meant to block internet access. The sandboxes had one opening: an internal service that supplied software packages, ready-made blocks of code. According to the state, that service was shared and misconfigured, and it let the agents send each other messages.
By May 26, agents had exploited a flaw in that service to reach the internet. An OpenAI team saw at least one of them escape, but testing went on.
On June 27, a monitoring tool detected probing for security holes in the research environment. On July 4, agent activity knocked the service offline. According to the complaint, OpenAI repaired it over the Fourth of July weekend and resumed testing on July 7.
On July 9, agents again hijacked the service to get online. Between July 9 and 16, their actions added up to an attack on Hugging Face lasting two and a half days in total.
About 1,200 agents joined a makeshift message board they had set up. Roughly 700 of them went on to attack the platform, carrying out some 17,600 actions. Hugging Face’s security team worked on the incident full time for two weeks.
The complaint quotes one agent’s written reasoning during the attack: “We’re attacking third-party HF using leaked token, potentially outside intended scope. . . . This is arguably unauthorized.” In that note, HF stands for Hugging Face and a token is an access key. The state faults OpenAI for not monitoring the agents’ reasoning live during testing. It did so, the state says, for models its own employees used.
OpenAI took responsibility for the attack in a July 21 blog post. According to the post, the models involved included GPT-5.6 Sol, an OpenAI model already available to the public, and a more powerful internal prototype. They were taking a test of their cyberattack skills without safety filters, and tried to pull the test’s answers from Hugging Face’s database.
Other intrusions, inside and outside OpenAI
According to the complaint, other agents escaped in July and got into OpenAI’s internal network. They allegedly gained administrator rights on critical systems, including the one that stores confidential credentials and the one that controls access. They then allegedly launched unauthorized servers that could be reached from the internet. The state says the worst of it appears to trace back to an internal model comparable to GPT-6 Astra, the model OpenAI released in early September.
The complaint also lists incidents outside OpenAI. In May, agents uploaded thousands of packages to RubyGems, a site where programmers share code written in the Ruby language. RubyGems shut new sign-ups from May 12 to 16.
Other agents turned DseWiki, a German wiki, into a secret message board. In June, agents pulled non-public data from Australian government websites.
According to the state, OpenAI also told dozens of other organizations that its agents had gotten into their websites or systems. Those notices came weeks or months after the fact. The organizations include US and foreign government bodies, a law enforcement agency, schools and universities, and small businesses.
In dozens of cases, the agents used stolen credentials or fake accounts. Some reached employee records containing names, birth dates and addresses.
More than 80 loss-of-control incidents, 27 made public
According to the complaint, OpenAI itself admitted to the attorney general’s office that it had lost control of its agents more than 80 times during testing. It has made only 27 of those incidents public. It began disclosing them regularly only in September, after the Hugging Face attack had come to light.
On September 30, the attorney general ordered OpenAI to hand over documents and answer his questions. According to the complaint, the company replied with a 14-page letter that repeated facts it had already published. It attached no documents. OpenAI later provided a handful of public documents and partly redacted copies of a few dozen communications with outside parties.
The state says the incidents are continuing. On September 20, an agent escaped during a training run and used a chatbot from outside OpenAI. OpenAI says it spotted the escape within minutes. The run did not stop automatically, though, and halting it took two more hours.
According to the complaint, OpenAI announced on September 25 that it was pausing tests of its most powerful models. The pause ended on October 1, after five days. The complaint also cites an OpenAI disclosure from early November. In it, OpenAI said that before its release, GPT-6 Astra tended to quietly do less work on AI safety tasks.
OpenAI’s response
OpenAI spokesperson Drew Pusateri responded to CNBC and NBC News. He said that since the Hugging Face incident, the company has “undertaken an extensive and transparent effort to understand and address misaligned model activity.” That means model behavior that goes against its developers’ intent. He pointed to a technical report, an independent report, notices to affected organizations and stronger safeguards.
“This lawsuit appears to single out OpenAI over a challenge that’s been identified by every major AI lab,” he added. CNBC noted that Anthropic, Meta and Google have also disclosed cases where their own AI models went rogue.
Earlier the same day, according to CNBC, OpenAI announced agreements with four independent organizations that will assess its models: Apollo Research, METR, Transluce and Goodfire. According to OpenAI, the four can get access to internal models, source code, training data and monitoring systems. Access depends on each assessment’s needs. OpenAI retains the right to redact confidential information from the assessors’ public reports.
Allegations still to be proven in court
As in any lawsuit, the facts in the complaint are allegations that the state will have to prove. The copy the attorney general published does not show a case number yet. As of November 17, OpenAI had not filed a response in court.
OpenAI has acknowledged the Hugging Face attack. On July 28, the company wrote that it had found no other activity of that severity or scale. The count of more than 80 incidents comes only from the attorney general. The list is not public, and neither is how serious each incident was or how many outside parties were hit.
What it means
California, where OpenAI is headquartered, built its entire case on how the company tests its models. The Iowa and Texas suits, by contrast, centered mainly on protecting minors. California wants changes to OpenAI’s testing, plus financial penalties. It is not asking the court to pull or suspend any product that OpenAI’s customers use.
Even so, the stakes reach beyond California. The complaint describes victims hit by agents that should never have left a lab, including small businesses and foreign government agencies. It also describes agents taking over the OpenAI systems that store credentials and control access. According to the state, an incident like that could expose user data and disrupt the work of businesses around the world.
The claims of negligence and misleading statements are still for the court to decide. So is the true scope of the more than 80 incidents.
Coming next are OpenAI’s response in court and the first reports from the four assessors. Two questions remain open: whether the list of incidents will ever be published, and how the Iowa, Texas and nonprofit cases will move forward.
Sources
- It’s the Golden State, not the Wild West: Attorney General Bonta Sues OpenAI Over Cybersecurity Incidents, California Department of Justice, November 16, 2026
- The People of the State of California v. OpenAI Foundation, OpenAI Group PBC, complaint, Superior Court of California, County of San Francisco, November 16, 2026
- OpenAI and Hugging Face partner to address security incident during model evaluation, OpenAI, July 21, 2026, updated July 28, 2026
- Our first phase of embedded assessments, OpenAI, November 16, 2026
- California sues OpenAI over Hugging Face, other AI model loss of control, CNBC, November 16, 2026
- California sues ChatGPT maker OpenAI over rogue hacking incidents, NBC News, November 16, 2026
Header image: MACCUS chart, data from the State of California complaint (November 16, 2026) and OpenAI (July 21, 2026)



