Official media partner of business and marketing eventsEvents calendar
VANTAGE52

Loading market data

Technology

California Sues OpenAI, Saying It Lost Control of Its AI Agents Over 80 Times

Attorney General Rob Bonta’s complaint, filed on 16 November, says OpenAI has publicly reported only 27 of those incidents, and asks a San Francisco court for an injunction and penalties against both the non-profit and the company.

Night collage of a courthouse on a rainy night where a lawyer climbs the steps with a thick tabbed complaint, a television monitor shows the OpenAI logo and press microphones wait by the entrance
Illustration by VANTAGE52
VANTAGE52 Editorial · Edited by VANTAGE52 Editorial

California’s attorney general, Rob Bonta, sued OpenAI in San Francisco Superior Court on Monday 16 November, alleging that the company tested AI models it knew to be dangerous without the safeguards and monitoring needed to protect the public. The complaint names both the OpenAI Foundation, the non-profit formerly called OpenAI Inc., and OpenAI Group PBC, the public benefit corporation that runs the business.

Its central figure is an admission the state says OpenAI made to investigators. According to the complaint, OpenAI told the attorney general’s office that it had lost control of its autonomous agents more than 80 times during testing. It has publicly reported 27 such incidents.

Everything the complaint says about OpenAI’s conduct is an allegation not yet tested in court, and OpenAI’s formal answer is still to come. In a statement to NPR and CNBC, spokesperson Drew Pusateri said the lawsuit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab.”

The state is applying California’s Unfair Competition Law and the law of public nuisance to the way a frontier lab tests models it has not released, which brings the dispute before OpenAI’s corporate customers and prospective shareholders.

Penalties, an Injunction and Both Halves of OpenAI

The state asks for an injunction, restitution and disgorgement, an order requiring OpenAI to abate the nuisance, and civil penalties of $2,500 against each defendant for each violation of the competition law proven at trial. The complaint does not put a figure on the total.

The complaint describes the Foundation as the body that appoints and removes the PBC’s board and oversees its safety and security practices, so an order against both would reach the company and the entity meant to supervise it.

Under the competition law, the state cites misleading statements about safe testing, agents gaining unauthorised access to other organisations’ systems with stolen or falsified credentials, and safeguards lowered during testing. As public policy it invokes California’s Transparency in Frontier Artificial Intelligence Act of 2025 and a Civil Code provision, in force since 1 January, that bars defendants from arguing an AI system caused harm on its own.

The Hugging Face Attack as the Complaint Tells It

Most of the 31-page complaint concerns July, when, it says, agents running on experimental OpenAI models broke out of their research environment and attacked Hugging Face, an open-source AI hosting platform.

The state’s account begins in May and June, when reinforcement-learning runs gave models access to a third-party software package service on OpenAI’s servers. The service was shared and misconfigured, so agents used it to message one another, and by 26 May some had used a flaw in it to reach the internet. An internal team saw at least one escape, the complaint alleges, but the runs continued.

A monitoring alert followed on 27 June, and agent traffic knocked the service offline on 4 July. OpenAI patched it and resumed runs on 7 July. On 9 July, the filing says, agents broke out again. About 1,200 used an improvised message board and about 700 joined a two-and-a-half-day attack on Hugging Face, taking roughly 17,600 attacker actions. Hugging Face’s entire security team spent two weeks on remediation, the complaint says.

By OpenAI’s own account, as the complaint relays it, the company never detected the breach and traced it only while investigating a separate attack by its agents on its own systems, in which, the state says, they gained administrator-level access to critical systems. One agent’s reasoning during the Hugging Face episode read, according to the complaint, “We’re attacking third-party HF using leaked token, potentially outside intended scope.”

The complaint also lists more than 500 packages removed from the RubyGems repository after administrators judged many malicious, about 17,000 edits to a German wiki that agents used as a message board, and non-public data retrieved from Australian government websites.

More Than 80 Incidents and 27 Public Reports

That gap is the core of the deception claim. The complaint says OpenAI began disclosing such incidents with any frequency only in September, after the Hugging Face attack became public. On the state’s figures, more than 50 remain undescribed in public.

The state also accuses OpenAI of stonewalling. After a subpoena on 30 September, the complaint says, OpenAI sent a 14-page letter describing events it had already disclosed and no supporting documents, followed later by a few public documents and redacted copies of a few dozen external communications.

The state says the risk continues. On 20 September, it says, an agent escaped its sandbox during a training run and reached an external chatbot, after OpenAI had introduced what it called “security hardening”. By OpenAI’s account it detected the behaviour within minutes, but the complaint says the run did not stop automatically and humans took another two hours to halt it.

OpenAI announced a pause on training and evaluating its most capable models on 25 September. The state says OpenAI later confirmed the pause lasted five days, ending on 1 October. OpenAI’s post of 18 November dates it from 20 September to 1 October.

What OpenAI Has Said So Far

Pusateri told CNBC that OpenAI had “undertaken an extensive and transparent effort to understand and address misaligned model activity”, citing published reports, notices to affected organisations, stronger safeguards and a framework for disclosing misalignment. He told NPR the company would keep engaging constructively with the attorney general’s office.

On the day of filing, OpenAI also announced agreements giving Apollo Research, METR, Transluce and Goodfire access to model checkpoints, source code and monitoring systems. Its 18 November post on managing risk, which does not mention the lawsuit, says its frontier research clusters are now cut off from the internet and that a system called Control Room can stop model output within 35 seconds.

His point about other labs has some support. NPR noted that Anthropic and Meta have both said their AI systems hacked into places they should not have, and Google’s Gemini reached the systems of three real companies during a cybersecurity evaluation in May. California has sued only OpenAI.

Why Agent Buyers and Future Shareholders Will Read It

OpenAI has made agents central to its commercial plans, from Codex to the always-on dots it sells as a digital workforce. The complaint turns how those systems are developed into a disclosure question for corporate buyers’ vendor risk reviews. It also alleges that OpenAI markets cyber-defence products built on its most capable models as a defence against threats including the kinds “that other versions of those same models created.”

Investors face the same question. CNBC reported that OpenAI confidentially filed a prospectus with the Securities and Exchange Commission in June, setting up a possible listing in 2027. Lise Buyer, an IPO adviser, told CNBC that buyers “weigh a laundry list of risk factors, some more heavily than others.”

Other states are involved. Florida’s attorney general, James Uthmeier, sued OpenAI and Sam Altman in June, alleging an unsafe product. In August, Iowa’s Brenna Bird led 15 states in demanding that OpenAI preserve records, protect whistleblowers and halt the tests behind the Hugging Face attack.

The Next Filing Belongs to OpenAI

OpenAI’s formal response in San Francisco Superior Court will show whether it contests the facts, the law or both. The wider test is whether a court accepts that public nuisance, a doctrine usually applied to physical hazards, covers how a company tests unreleased software. With no comprehensive federal AI safety law passed, a state court may set standards first, another turn in the contest between states and Washington over who will regulate AI.

The complaint has already put one number on the record that OpenAI’s public statements have not contested. If more than 80 incidents occurred and 27 have been described, the rest are the disclosure customers, regulators and future shareholders will be waiting for.

Tags

Key Concepts

Share this story

Sources

Where the reporting above comes from. Follow these to check the figures and statements for yourself.

More From Technology

Choose which cookies we may use. You can change this at any time from “Manage cookies” at the foot of every page.