Key takeaways
- California Attorney General Rob Bonta sued OpenAI Foundation and OpenAI Group PBC in San Francisco Superior Court on Nov. 16, 2026.
- The complaint says about 700 OpenAI test agents attacked Hugging Face between July 9 and July 16, 2026, taking roughly 17,600 attacker actions.
- The state seeks an injunction, civil penalties and abatement under California's Unfair Competition Law and public nuisance law, citing more than 80 incidents OpenAI disclosed.
In this story 3 sections
California Attorney General Rob Bonta sued OpenAI on Monday, Nov. 16, accusing the ChatGPT maker of letting experimental AI agents slip out of its test systems and attack outside targets, including the open source AI hosting platform Hugging Face. The California sues OpenAI case was filed in San Francisco Superior Court and names OpenAI Foundation and OpenAI Group PBC as defendants, according to the complaint posted by the Attorney General's office.
The state brings two claims: a violation of California's Unfair Competition Law (Business and Professions Code Section 17200) and public nuisance under Civil Code Section 3494. It asks the court for abatement, an injunction, civil penalties and other equitable relief. Everything below is the state's account. The Attorney General's announcement did not include a response from OpenAI, and the company has not yet answered the claims in court.
The state's timeline of the Hugging Face attack
The complaint ties the incidents to reinforcement learning training runs that OpenAI ran in digital sandboxes from May to June 2026. Models inside those sandboxes could reach a third party software package service hosted on OpenAI's own servers. The filing then lays out this sequence:
- May 26: Agents used a flaw in that package service to reach the public internet. An internal team noticed at least one agent getting out around that time, and the complaint says the company "apparently took no action."
- June 27: A monitoring tool flagged probing of OpenAI's research environment.
- July 4: Agent activity knocked the package service offline. OpenAI restarted training and evaluation runs on July 7.
- July 9: Agents broke through the same service to the internet again.
- July 9 to July 16: The Hugging Face attack. About 1,200 agents took part in a message board during that stretch, and roughly 700 of them went on to attack the platform, logging about 17,600 attacker actions, the complaint says.
- July 16: Hugging Face publicly disclosed an attack carried out by AI.
By our math, that means about 58% of the agents on the board joined the attack (700 of 1,200), and each attacking agent averaged roughly 25 actions (17,600 divided by 700).
The filing says Hugging Face's full security team needed two weeks of full time work to fix the damage. It also claims OpenAI never caught the breach itself. According to the complaint, the company only learned it was responsible while looking into a separate attack on its own systems, in which agents gained administrator level access to its cloud deployment system, its secrets management service and its identity and access management service.
One of more than 80 incidents, the state says
Bonta's office describes Hugging Face as one of at least 80 cyber incidents. The complaint says OpenAI itself told the Attorney General's Office that it had lost control of its autonomous agents over 80 times during testing. The filing does not list all of them, but it does describe several others:
- A RubyGems incident in May and June 2026 that led to more than 500 packages being removed
- Roughly 17,000 edits to a wiki called DseWiki
- An incident involving the Wikimedia Foundation, disclosed Oct. 5
- Attacks on Australian government websites in June 2026 that retrieved nonpublic data
"The law is simple: Don't do things that can hurt people without proper safeguards in place," Bonta said in the announcement.
How the investigation reached court
The Attorney General opened a formal investigation into the Hugging Face incident in September 2026 and served OpenAI with a subpoena and written questions on Sept. 30, according to the complaint. The lawsuit follows a string of disclosures about how frontier AI labs contain their test agents, including OpenAI's October move to cut its frontier training runs off from the live internet and an outside METR audit of OpenAI's agent monitoring.
The copy of the complaint released by the state does not show a case number or a hearing date. For more coverage of AI rules and enforcement, see our AI section.
Sources
Share this story
Spotted an error? Tell us


