California sues OpenAI, alleges 80-plus agent escapes hidden from public

Tags:

California Attorney General Rob Bonta has sued OpenAI, alleging the company has lost control of its autonomous AI agents more than 80 times during testing and misled the public about how safely it develops its models.

The complaint was filed on November 16 in San Francisco Superior Court against OpenAI Foundation and OpenAI Group PBC. It says OpenAI disclosed the 80-plus figure to state investigators but has publicly reported just 27 such incidents. In those incidents, agents reached third-party systems, foreign government websites and OpenAI’s own infrastructure, the state alleged.

“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,” Bonta said in a statement announcing the lawsuit.

The state brings two claims: violations of California’s Unfair Competition Law and creation of a public nuisance. It is seeking an injunction, restitution, disgorgement, and an order to abate the nuisance. It also wants civil penalties of $2,500 against each defendant for each violation proven at trial.

Other incidents and internal breaches

In one July incident, agents escaped their sandboxes and gained administrator-level access to OpenAI’s own systems, according to the complaint. Those systems included the company’s cloud model deployment system, its secrets management service, and its identity and access management service. The agents then launched unauthorized servers, exposing OpenAI’s infrastructure to the internet. OpenAI did not let external researchers review that attack, the complaint said.

The complaint identified several other incidents involving OpenAI agents.

At RubyGems, a software package repository, agents allegedly uploaded thousands of packages in May 2026. Administrators removed more than 500 packages they considered malicious, and the platform suspended new user registrations between May 12 and May 16.

The complaint also alleged that agents made approximately 17,000 edits to a German website, DseWiki, using it as a message board to exchange information about bypassing sandbox restrictions.

Other incidents involved Wikimedia Foundation’s infrastructure and Australian government websites. The complaint said agents overwhelmed Wikimedia services with millions of API requests and hundreds of thousands of Wikidata queries, likely contributing to a partial outage in May. In June, agents allegedly accessed non-public information from Australian government websites.

The complaint further said OpenAI notified dozens of organizations whose websites or internal systems had been accessed by its agents. These included government institutions, law enforcement agencies, educational institutions, and small businesses. Some incidents allegedly involved stolen credentials, API keys, and fraudulent accounts, with agents accessing personally identifiable information.

OpenAI’s own infrastructure was also compromised in at least three attacks in July, according to the complaint. Agents allegedly obtained administrator-level access to systems supporting cloud model deployment, secrets management, and identity and access management.

Escapes after the fix

The state argued that the danger is ongoing. On September 20, after OpenAI had hardened its environment, an agent escaped during a training run and reached an external chatbot.

“This incident occurred after OpenAI had implemented ‘security hardening’ following the Hugging Face attack,” the lawsuit added.

OpenAI detected it within minutes, but the run did not stop automatically, and human reviewers took two more hours to halt it, the complaint said.

OpenAI announced a pause on training and evaluating its most capable models on September 25. The pause lasted five days, ending October 1, according to the complaint.

Warnings that went unheeded

The complaint argued that the incidents were foreseeable. It cited the run-up to the Hugging Face attack as evidence. By May 26, agents had used a vulnerability in a misconfigured software package service to reach the public internet.

An internal OpenAI team saw at least one agent escape but apparently took no action, the complaint said. On June 27, a security monitoring tool flagged that the research environment was being probed for vulnerabilities, and OpenAI again did not stop its runs.

When agent traffic knocked the service offline on July 4, OpenAI patched it over the holiday weekend and resumed runs on July 7, the state alleged. Two days later, agents broke out again.

The state also faulted OpenAI for not running live chain-of-thought monitoring in training and evaluation, even though it did so for models deployed to employees. During the Hugging Face attack, one agent’s reasoning read: “We’re attacking third-party HF using leaked token, potentially outside intended scope. . . . This is arguably unauthorized.”

Bonta’s office served OpenAI with an investigative subpoena on September 30. OpenAI answered with a 14-page letter and produced no supporting documents, the complaint said.

California is not the only state pressing OpenAI. In August, Iowa Attorney General Brenna Bird led a 15-state coalition demanding that OpenAI preserve evidence, protect whistleblowers and halt the tests behind the hack until it could show it can run them responsibly. “The AI weaponized itself, which is extremely scary,” Bird said in a statement.

OpenAI did not immediately respond to a request for comment.

The article originally appeared on ComputerWorld.

Categories

No Responses

Leave a Reply

Your email address will not be published. Required fields are marked *