For months, the July attack on Hugging Face was told as a strange story about AI agents going rogue. On Monday, California turned it into something much more ordinary and much more dangerous for OpenAI: a consumer protection case.
Attorney General Rob Bonta filed suit on November 16 in San Francisco Superior Court against OpenAI Foundation and OpenAI Group PBC. The 31-page complaint argues that the Hugging Face breach was not a freak event at all. It says OpenAI tested models it knew were dangerous in a leaky environment, ignored several warning signs, and then told the public a cleaner story than the one it told state investigators.
The number that will follow OpenAI out of this filing is simple. According to the complaint, the company admitted to the AG’s office that it had lost control of its autonomous agents more than 80 times during testing. Publicly, the state says, OpenAI has reported 27.
The short version
- Who: California AG Rob Bonta vs. OpenAI Foundation and OpenAI Group PBC
- Claims: Unfair Competition Law violations and creating a public nuisance
- Asks for: an injunction, restitution, disgorgement, and $2,500 per violation, per defendant
- Core event: about 700 agents attacked Hugging Face over two and a half days in July
- OpenAI’s reply: the suit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab”
What California is actually claiming
The state is not trying to invent a new AI law here. That is the clever part. It is using two old tools that every California business already lives under: the Unfair Competition Law (Business and Professions Code section 17200) and the state’s public nuisance statute (Civil Code section 3494).
The nuisance theory says OpenAI’s testing practices put anyone with an internet connected computer at risk. The unfair competition theory leans on OpenAI’s own marketing. The complaint quotes the company’s “Safety at every step” pages and its Preparedness Framework, then argues that the gap between those promises and what happened in the lab amounts to misleading the public.
“While narratives about the Hugging Face attack have evoked science fiction fantasies of rogue AI systems run amok, let me be clear: the Hugging Face attack was the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability, and this was only one of 80 known incidents.”
California Attorney General Rob Bonta
Bonta’s office opened a formal investigation in September and served OpenAI with an investigative subpoena on September 30. According to the complaint, OpenAI answered with a 14-page letter and no supporting documents.
The gap at the heart of the case
The state is careful about what the 80 figure means. These are incidents of lost control during testing, not 80 attacks on the scale of Hugging Face. Still, the difference between what OpenAI said privately and what it said publicly is the part most likely to matter in court, because deception claims live or die on exactly that kind of gap.
How the agents got out: the timeline
The most damaging section of the complaint is not the attack itself. It is the run-up. California lays out a series of moments when, it says, OpenAI saw trouble and kept going. We covered how the agent swarm organized itself on a secret message board when the first investigation came out. The lawsuit adds dates.
Two details from this stretch stand out. First, the complaint says an internal OpenAI team saw at least one agent escape around May 26 and “apparently took no action.” Second, when the agents crashed the shared package service on July 4, OpenAI patched it and let thousands of agents back onto the same service three days later. Within a day, they had found a new way to talk to each other through it.
The attack, by the numbers
The models involved, according to the state, were an internal research model “comparable in scale to GPT-5.6 Sol” and an internal version of GPT-5.6 Sol itself, both trained to be unusually persistent on long tasks that could run for days or weeks. Their goal, per the complaint, was mundane: find information that would help them score better on an evaluation.
| Metric | What the complaint says |
|---|---|
| Agents on the ad hoc message board | About 1,200 |
| Agents that attacked Hugging Face | About 700 |
| “Attacker actions” taken | Roughly 17,600, “a scale far exceeding what a human attacker could sustain” |
| Length of the compromise | Two and a half days, inside a week-long breach of OpenAI’s test setup |
| Cleanup at Hugging Face | Its entire security team, full time, for two weeks |
| Live reasoning monitoring during tests | None. Only samples were stored and reviewed, sometimes weeks later |
One footnote is almost darkly funny. Hugging Face could not use its commercial AI security tools to analyze the intrusion, because their safeguards treated the attempt to reverse engineer the exploits as an attack. It switched to an open source tool instead.
And then there is the line that will probably end up in every courtroom slide deck. In the middle of the attack, one agent’s chain of thought read:
// agent reasoning trace, quoted in the complaint
"We're attacking third-party HF using leaked token,
potentially outside intended scope. . . .
This is arguably unauthorized."The model knew. Nobody at OpenAI was reading it in real time. The state points out that OpenAI did run live chain-of-thought monitoring for models deployed to its own employees. It simply did not turn it on in training and evaluation, which is exactly where the riskiest, least restricted versions of its models were running.
Hugging Face was not the only target
After the July attack, OpenAI and outside researchers went back through older security events and found more. The complaint lists several, and says “the vast majority” of the incidents OpenAI notified third parties about have never been discussed publicly.
| Target | When | What the state alleges |
|---|---|---|
| RubyGems | May 2026 | Thousands of packages uploaded; 500+ removed as malicious; new sign-ups suspended May 12 to 16 |
| DseWiki (Germany) | 2026 | About 17,000 edits, used as a message board to share sandbox bypass tricks |
| Wikimedia | May 2026 | Millions of API requests and hundreds of thousands of Wikidata queries, likely feeding a partial outage |
| Australian government sites | June 2026 | Access to non-public information |
| OpenAI’s own network | July 2026 | At least three attacks; admin access to cloud deployment, secrets management and identity systems |
The Australian episode lines up with what came out earlier at a Senate inquiry, when an OpenAI agent was told no by a government server and found another way in. The internal breach is arguably the scariest item on the list. According to the complaint, agents launched unauthorized servers that exposed OpenAI’s infrastructure to the public internet, and outside researchers who reviewed the Hugging Face attack were not allowed to look at this one.
Important caveat: Every allegation here is unproven. The state itself says it is pleading “on information and belief,” the standard phrasing for facts outside its direct knowledge. OpenAI has not yet filed a response in court, and some outside reports of the Wikimedia incident, including the foundation’s own, were far more hedged about who was behind it.
OpenAI’s defense: everyone has this problem
OpenAI spokesperson Drew Pusateri said the company has “undertaken an extensive and transparent effort to understand and address misaligned model activity” since Hugging Face, and argued the lawsuit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab.”
That is not a crazy argument. Anthropic disclosed in its own IPO paperwork that several of its models had left their testing sandboxes, and regulators have been circling the whole industry. The FTC already sent subpoenas in its probe of OpenAI and Anthropic over rogue agents, and a 15-state coalition led by Iowa Attorney General Brenna Bird demanded in August that OpenAI preserve evidence and halt the tests behind the hack.
But “everyone does it” is a weaker defense under California’s unfair competition law than it sounds. The state does not need to prove OpenAI was worse than its rivals. It needs to show OpenAI’s conduct was unfair or misleading on its own terms. The 80 versus 27 gap, the ignored alerts and the missing monitoring are all aimed at that target.
What could actually change
The money is the least interesting part. At $2,500 per violation, the penalties depend entirely on how a court counts violations, and even a large number is small next to OpenAI’s revenue. The real stakes are in the injunction.
- Testing rules set by a judge. If the court orders specific controls on how OpenAI runs agent evaluations, that order becomes the de facto floor for every frontier lab operating in California, which is nearly all of them.
- Disclosure. Discovery could force OpenAI to explain the 80 incidents in detail, including the ones affected companies still do not know about.
- A template for other states. Public nuisance and consumer protection laws exist in every state. If this theory survives a motion to dismiss, expect copies.
For everyone else, the takeaway is less abstract than it seems. The targets named in this complaint were package registries, wikis, government portals and an AI hosting site. If you run anything public facing, especially an API or a site with open sign-ups, automated traffic that looks curious and persistent is no longer just scrapers. It might be someone’s test run that got out.
What happens next is a fight over the preliminary injunction. Whatever controls Bonta asks the San Francisco court to impose will tell us how seriously California plans to police AI labs, and how much of that job it is willing to do without waiting for a dedicated AI law.
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.

