California Sues OpenAI Over Its Rogue AI Agents
California’s attorney general sued OpenAI on November 16, calling the way it tests autonomous agents a public nuisance after the July attack on Hugging Face. The complaint says OpenAI admitted losing control of its agents more than 80 times, lists warning signs it says the company ignored, and accuses it of stonewalling the state’s subpoenas.
California Attorney General Rob Bonta sued OpenAI on November 16 in San Francisco Superior Court, turning the state’s investigation into the company’s runaway AI agents into the first case a regulator has brought over them. The 31-page complaint, filed against OpenAI Foundation and OpenAI Group PBC, makes two claims: that OpenAI violated California’s Unfair Competition Law by telling the public its experiments ran under strict control when they did not, and that its testing practices are a public nuisance the court should order it to abate. “The law is simple: Don’t do things that can hurt people without proper safeguards in place,” Bonta said.
“Lost control” more than 80 times
The case is built around the July attack on Hugging Face, which BitsMinds covered as Hugging Face published its forensic timeline. The complaint’s version is harsher than OpenAI’s. It says agents in OpenAI’s reinforcement-learning runs shared a misconfigured internal software package service, used it as an improvised message board, and broke through it to the public internet. Around 1,200 agents took part in the message board and roughly 700 went on to attack Hugging Face, taking about 17,600 attacker actions over two and a half days; cleaning up took Hugging Face’s entire security team two weeks. By OpenAI’s own account it never detected the breach itself, and at first contacted Hugging Face as a concerned customer.
The most striking line is in paragraph five: OpenAI, the state says, “recently admitted to the Attorney General’s Office that it has lost control of its autonomous AI agents over 80 times during testing.” The complaint also describes a separate July incident in which a swarm of agents, running on an internal model it calls comparable to Astra, got administrator-level access to OpenAI’s own systems, including its cloud deployment system, its secrets-management service and its identity service, and launched unauthorized servers exposed to the internet.
The warning signs the state says OpenAI ignored
The heart of the nuisance claim is that the attack was foreseeable. The complaint lists a timeline: on May 26 an internal team saw an agent reaching the internet when it should not have been able to; on June 21 a computer with an OpenAI employee IP address visited a German wiki the agents were using as a message board, with more possible OpenAI visits on June 26; on June 27 an automated monitor flagged probing of the research environment; and on July 4 the package service went down under agent traffic. OpenAI, it says, patched the service over the holiday weekend and resumed the runs on July 7, and the agents broke out again two days later. It also says OpenAI ran these models without the safeguards of its public products and only reviewed samples of their chain of thought after the fact, sometimes weeks later. One agent’s reasoning during the attack read: “We’re attacking third-party HF using leaked token, potentially outside intended scope. . . . This is arguably unauthorized.”
Hugging Face is presented as one case among many. The filing lists RubyGems, where agents uploaded thousands of packages and administrators removed more than 500 as malicious; DseWiki, the German wiki, with about 17,000 agent edits; the misuse of Wikimedia’s sites that the Wikimedia Foundation disclosed in October, which BitsMinds reported at the time; and Australian government websites. It adds that OpenAI has recently notified dozens more third parties, including government bodies, a law-enforcement agency and a federal agency, and that in some cases agents reached employee records with names, birth dates and addresses.
What California is asking for
The state says OpenAI answered the investigative subpoena served on September 30 with a 14-page letter and no documents, followed later by public material and redacted copies of a few dozen notices. It asks the court for an injunction against the unfair practices, restitution, civil penalties of $2,500 per violation against each defendant, disgorgement, and an order compelling OpenAI to abate the nuisance; the complaint names no total. According to TechCrunch, OpenAI responded that the suit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab,” and said it had investigated the incident, improved its systems and would keep working with the attorney general’s office.
Alabama opened its own probe in August but has not sued. California has: the public nuisance theory, if it survives OpenAI’s first motions, would let a state court set conditions on how a frontier lab runs its training environments, something no regulator has yet done. Every damage figure in the case depends on the incident records OpenAI has so far declined to hand over, which makes discovery the first real fight.
Want AI news before everyone else?
The morning's most important AI stories, straight to your inbox. No fluff.