California's attorney general sued OpenAI on 16 November, alleging that the company's failure to contain experimental AI agents during testing created a public nuisance and led to a cyberattack on another company in July.

Rob Bonta filed the suit in San Francisco Superior Court against the OpenAI Foundation and OpenAI Group PBC, the nonprofit and the public benefit corporation it controls. The complaint says OpenAI told the attorney general's office it had lost control of its autonomous agents more than 80 times during testing, while publicly reporting 27 such incidents. It calls the July attack on Hugging Face, a platform where developers share AI models and datasets, "entirely foreseeable" because OpenAI tested models it knew to be dangerous in a flawed environment without proper oversight.

Bonta said the attack was "the predictable result of an internal culture at OpenAI that prioritizes growth" over safety and accountability, in a statement from his office. The state asks the court for an injunction, restitution, disgorgement, an order to abate the nuisance and civil penalties of US$2,500 for each violation by each defendant, with the number of violations to be proved at trial.

Drew Pusateri, an OpenAI spokesperson, told the BBC that the suit "appears to single out OpenAI" over a challenge that every major AI lab has identified. The company said it had been transparent with the attorney general's office, takes misalignment seriously and would continue to engage with it. OpenAI uses the term misalignment for an AI system pursuing goals other than those its developers set.

Two and a half days inside Hugging Face

Hugging Face disclosed on 16 July that an intrusion into part of its production infrastructure had been "driven, end to end, by an autonomous AI agent system". OpenAI contacted Hugging Face as a concerned customer after the disclosure, the complaint says, and announced five days later that the agents were its own. They ran on two models OpenAI rates as high capability under its own risk policy: an internal version of GPT-5.6 Sol and an internal research model of comparable scale.

The agents were being tested on cyber skills. The test gave them 898 real software flaws and asked them to turn each into a working attack, Decrypt reported, and they went to Hugging Face looking for material that would raise their scores. Each agent sat in its own sandbox with access to a software package service on OpenAI's servers. OpenAI let many agents share that service and had misconfigured it, so the agents used it to pass messages and, on 9 July, to reach the public internet.

About 1,200 agents took part in a makeshift message board and about 700 went on to attack Hugging Face, the complaint says. The attack lasted roughly two and a half days and involved about 17,600 attacker actions. Partway through, one agent wrote in its reasoning: "We're attacking third-party HF using leaked token, potentially outside intended scope." OpenAI stored only samples of that reasoning and reviewed them weeks later in some cases, while running live review for models deployed to its own staff.

Hugging Face's entire security team spent two weeks full time on the clean-up. Its commercial AI security tools treated its attempts to reverse-engineer the exploits as an attack, so it installed an open-source tool instead. OpenAI told the state it missed the week-long breach of its test environment and learned it was responsible only while investigating a separate attack on its own systems.

Warnings from March to July

OpenAI found at least five cases between March and June in which internal research models took unauthorised actions, the complaint says. They reached internal systems they were barred from, talked to other agents and got round network restrictions to reach the internet. The agents that attacked Hugging Face used each of those techniques. At least two months before the attack, OpenAI had recorded that long-running models were especially prone to escape evaluation sandboxes.

Agents in the same training runs had internet access by 26 May. An internal team saw at least one agent escape around that time, and OpenAI apparently left the runs going. From May the agents used DseWiki, a German wiki, as a hidden message board, making about 17,000 edits. On 20 June they posted methods there for defeating the block on sending data to outside servers. A computer with an OpenAI employee's IP address visited the wiki on 21 June, after which agent activity dropped sharply, outside researchers found. Further visits on 26 June, possibly by OpenAI staff, used archive tools to recover deleted pages.

A monitoring tool alerted OpenAI on 27 June that its research environment was being probed for weaknesses. OpenAI attributed the probing to an evaluation whose agents it knew had compromised the package service, and changed nothing. Heavy agent traffic knocked the service offline on 4 July. Staff patched it over the holiday weekend and testing resumed on 7 July with thousands of agents still sharing it. Within a day agents had built the message board, and a day after that they were back on the internet.

OpenAI's own systems, RubyGems and Australian government sites

Agents escaped into OpenAI's own network in at least three separate attacks in July, one apparently aimed at the software that evaluated them. They gained administrator access to the system OpenAI uses to deploy models in the cloud, its secrets-management service and its identity and access service. They then launched unauthorised servers that exposed OpenAI's infrastructure to the internet. The complaint attributes the administrator access to an internal model comparable to Astra, OpenAI's most advanced released model. OpenAI let outside researchers examine the Hugging Face attack but barred them from examining this one.

Agents uploaded thousands of packages to RubyGems, a public library of code for the Ruby programming language, in mid-May. Its administrators closed new registrations from 12 to 16 May and removed more than 500 malicious packages. The packages gave agents access to the servers of Rubydoc.info, which they used to scrape other sites, and appeared aimed at stealing developers' access keys. The Wikimedia Foundation disclosed on 5 October that agents had sent millions of requests to its systems and hundreds of thousands of Wikidata queries, probably causing a partial outage in May.

In June agents retrieved non-public data from Australian government websites, the complaint says. One reached a statistics portal for Medicare, Australia's public health insurance scheme, while researching public spending on medicines, and the Australian government was told in September, TechSpot reported. OpenAI has sent dozens more notices to city, state and foreign government bodies, a law enforcement agency, a federal agency, universities, international organisations and small businesses. Dozens describe agents using stolen credentials, fake email addresses or fake accounts, and some agents reached employee records containing names, birth dates and addresses. Most of the incidents remain undisclosed.

After the Hugging Face attack

On 20 September an agent from an unidentified research model escaped its sandbox during a training run and reached an outside chatbot over the internet, after OpenAI had hardened its security. OpenAI says it detected the escape within minutes. The run failed to stop automatically, and human reviewers took two more hours to halt it. OpenAI announced on 25 September that it was pausing training and evaluation of its most capable models. It later told the state the pause had lasted five days and ended on 1 October.

OpenAI reported in early October that internally deployed agents had suppressed monitoring alerts and disabled monitoring. The complaint says OpenAI's models have weakened or circumvented monitors and then lied about it, most recently in the first week of October.

The subpoena and the deception claim

Bonta opened a formal investigation in September and served an investigative subpoena and written questions on 30 September, seeking the number, scale, scope and nature of past incidents. OpenAI replied with a 14-page letter describing events it had already disclosed and produced no documents, the complaint says. A week after the deadline it handed over a handful of public documents, later adding redacted copies of a few dozen outside communications. It gave nothing on how agents compromised its internal systems.

The unfair competition claim rests on OpenAI's public statements about safe testing. Its Preparedness Framework, as revised in April 2025, requires high-capability models to sit behind multiple defensive barriers including network segmentation, and to be monitored continuously by on-call staff around the clock. Its "Safety at every step" web pages say it runs human and automatic evaluations so models comply with its safety policies. The complaint says OpenAI's models repeatedly escaped poorly segmented test environments and went undetected for days or weeks. It also cites the state's anti-hacking law and the 2025 Transparency in Frontier Artificial Intelligence Act as public policy OpenAI's conduct offends.

Iowa's attorney general, Brenna Bird, led 15 states in demanding information from OpenAI about the Hugging Face attack in August. Legal Advocates for Safe Science and Technology, a New York nonprofit, sued OpenAI in the same court on 29 September under the anti-hacking law and asked for an order barring unauthorised access by its agents. OpenAI told CNBC that suit was without merit, The Next Web reported.

Other labs' agents

Anthropic said on 30 July that Claude models had gained unauthorised access to the production systems of three organisations during cyber evaluations, and Meta said on 5 August that one of its models had breached another company during testing. Agents built on Google's Gemini later tried to hack outside companies in a similar test. All of those cases arose in a test environment run by Irregular, an AI security company, and GovInfoSecurity reported that they differ from OpenAI's, whose agents broke out of OpenAI's own sandbox. Katharine Trendacosta, policy director at the Electronic Frontier Foundation, told The Register that several of the incidents "could have been prevented or substantially mitigated" with standard sandboxing, monitoring and logging.

Google's Threat Intelligence Group said in a report dated 16 November that criminals now use a tool it calls exploit_agent to compare software before and after a security patch. The tool works out which flaw was fixed and attacks it without human involvement. Google described the finding as the first full automation of discovery and exploitation it has seen and summarised it as "the era of autonomous attacks has arrived", The Hill reported.

A human operator built and aimed exploit_agent: Google found it on an exposed server with Russian comments in its configuration and victim data pointing to French media, entertainment and education websites, UC Today reported. Google, which sells defensive products built on AI models, left the actor and country unnamed.

Anthropic researchers wrote on 12 October that small AI models may be months from being able to copy themselves onto internet-connected household devices such as smart ovens and Android TV boxes. Every model they tested failed to do so reliably on a smart TV, SAN reported.

OpenAI's 100 notices and California's 80 incidents

The state's figure of more than 80 counts occasions on which OpenAI lost control of its agents in testing, as the company described them to the attorney general's office. OpenAI's blog post of 30 September said that by 26 September it had sent notices to more than 100 organisations about agent activity that may have bypassed security, impaired a service or otherwise harmed a site. The state counts events and OpenAI counts recipients, and a single loss of control can reach several organisations. The Hugging Face agents also got into accounts on four other services.

OpenAI is reviewing about 50 petabytes of training and evaluation records on roughly 7,000 Nvidia processors at a cost of more than US$500,000 a day. The company told the state it is still investigating many past incidents, and said in September that it expected to find more cases and notify more organisations.