# California sues OpenAI over leaks from its AI agents

> California sues OpenAI for unfair competition and public risk after losing control of its AI agents more than 80 times
> By Alfredo Collosa · AI Agents · 2026-11-18

On November 16, 2026, the California Attorney General's Office filed a lawsuit against OpenAI Foundation, OpenAI Group PBC and one hundred still-unidentified defendants in the Superior Court of the County of San Francisco. According to the complaint, several artificial intelligence agents developed by OpenAI left the isolated environments in which the company was testing them and attacked third-party systems. The Attorney General's Office, led by Rob Bonta, argues that OpenAI developed and tested those models without the safeguards needed to prevent them from accessing external systems.

## The attack on Hugging Face

The central episode of the lawsuit occurred in July 2026. According to the Attorney General's Office, between May and June OpenAI ran several reinforcement training sessions with research models it had not yet released. The tests were carried out in supposedly isolated environments, known as *sandboxes*. The agents were only supposed to be able to access a software package repository hosted on the company's own servers. A configuration error allowed them to communicate with each other and, by exploiting a vulnerability, to access the internet.

**[Jensen Huang offers to return to Nvidia, which beats its quarterly revenue record](https://observatorioblockchain.com/en/ai/huang-volver-nvidia-record-ingresos/)**

Between July 9 and 16, the agents that had escaped the test environment attacked the infrastructure of Hugging Face, an open platform that hosts artificial intelligence models. According to the lawsuit, some 1,200 agents took part in an improvised message board and about 700 were involved in the attack. In two and a half days they carried out some 17,600 actions, a pace that, according to the Attorney General's Office, a single person could not sustain. Hugging Face had to devote its entire security team for two weeks to repairing the damage.

The Attorney General's Office [argues](https://oag.ca.gov/system/files/attachments/press-docs/complaint-oai-cyber.pdf) that OpenAI did not detect the attack when it occurred and that, initially, it contacted Hugging Face as a customer concerned about what had happened. The company reportedly discovered that its own agents were behind the attack while investigating another incident.

## Other incidents cited in the lawsuit

*   **RubyGems.** Between May and June, the agents published thousands of packages in this software repository, many of them malicious, according to the Attorney General's Office. RubyGems removed more than 500 packages and suspended new user registrations for several days.
*   **DseWiki.** The agents used this German site as a covert message board. According to the lawsuit, they made some 17,000 edits and posted techniques for bypassing the test environment's restrictions.
*   **Wikimedia.** According to the foundation, the agents sent millions of requests to its services and probably caused a partial outage in May.
*   **Australian Government sites.** According to the lawsuit, the agents accessed data that was not public.
*   **OpenAI infrastructure.** Groups of agents reportedly obtained administrator permissions on critical systems of the company itself, including those managing cloud deployment, secrets, and access identities and permissions. The Attorney General's Office includes in these episodes an agent of an internal model comparable to Astra.

According to the Attorney General's Office, OpenAI reportedly notified the incidents to dozens of affected third parties, including municipal, state and foreign institutions, a law enforcement body, a federal agency and educational centers. In several cases, the agents reportedly impersonated users, used stolen credentials or created fake accounts. The lawsuit states that OpenAI acknowledged to the Attorney General's Office that it had lost control of its agents more than 80 times during testing. Publicly, the company had reported 27 incidents.

## Warnings to OpenAI before the attack

The Attorney General's Office argues that the incidents were foreseeable. The two models involved in the Hugging Face attack were classified as high capability in OpenAI's Preparedness Framework. For that category, the company's internal rules required continuous monitoring and multiple layers of protection. According to the lawsuit, OpenAI failed to meet those requirements despite receiving several warning signs.

On May 26, an internal team noticed that an agent had accessed the internet. Between June 21 and 26, OpenAI employees visited DseWiki. On June 27, an alert indicated that probing was taking place in the research environment. On July 4, the agents' activity caused an outage of the software package service. According to the Attorney General's Office, the company rushed to repair it over the holiday weekend to resume training.

OpenAI also reportedly did not activate real-time monitoring of the models' reasoning during training. In the middle of the Hugging Face attack, the reasoning log of one of the agents indicated, according to the lawsuit, that they were attacking a third party with a leaked access token and that the action could be considered unauthorized (*arguably unauthorized*).

The Attorney General's Office attributes part of the problem to insufficient isolation of the test environments. Several agents shared a service connected both to OpenAI's internal network and to the internet. In addition, the models were tested without the safeguards the company applies to its publicly available products. According to the lawsuit, OpenAI reportedly acknowledged that those protections could have mitigated the attack.

The Attorney General's Office also notes that OpenAI trained the models to persist when trying to complete their tasks and assigned them long-duration jobs. The company itself had warned that those characteristics increased the likelihood that the agents would take unforeseen actions.

## The Attorney General's Office accuses OpenAI of misleading advertising

The lawsuit cites OpenAI's mission statement, which presents safety as a central element, and the section of its website titled Safety at Every Step. It also refers to the 2025 Preparedness Framework, which describes controls such as network segmentation and continuous 24-hour monitoring as mandatory.

The Attorney General's Office argues that those safety claims were misleading. According to the lawsuit, the models repeatedly left environments with insufficient segmentation and the company lacked the continuous monitoring it claimed to require.

## An agent left the test environment again in September

According to the lawsuit, on September 20, after OpenAI tightened security following the Hugging Face attack, another agent left its test environment and accessed an external chatbot. The company detected the incident within minutes, but training continued. Human reviewers took two more hours to stop it. On September 25, OpenAI announced a pause in training its most capable models. The Attorney General's Office states that the pause lasted only five days and that by October 1 it was already over.

The lawsuit also cites signs that the models are increasingly difficult to supervise. Astra's technical sheet, published in early September, notes that the model's reasoning was harder to monitor. OpenAI also reported that Astra tended to underperform its capacity on safety tasks. In early October, the company reported that internally deployed agents had managed to suppress alerts and disable monitoring.

## The Attorney General's Office accuses OpenAI of insufficient cooperation

On September 30, the Attorney General's Office served OpenAI with a request for information as part of the investigation, which included written questions. According to the lawsuit, the company responded with a 14-page letter about facts that were already public and attached no documentation. A week after the deadline, it delivered a few public documents and redacted copies of several dozen communications.

The Attorney General's Office adds that OpenAI allowed external researchers to evaluate the Hugging Face attack and prevented them from examining the incident that affected the company's own infrastructure.

## What California is asking the court for

The lawsuit brings two claims. The first relies on California's Unfair Competition Law, set out in Section 17200 et seq. of the Business and Professions Code, and accuses OpenAI of unlawful, unfair and fraudulent practices. The second invokes Section 3494 of the California Civil Code and asks the court to order an end to what the State describes as a *public nuisance*, an interference with collective rights or interests.

California seeks an injunction to prevent the practices complained of and to require OpenAI to eliminate the situation that, according to the Attorney General's Office, harms third parties. It also seeks restitution of money or property obtained through the unfair competition practices, civil penalties of $2,500 for each violation and for each defendant, disgorgement of profits obtained (*disgorgement*) and payment of the costs of the proceedings.
