California Attorney General Rob Bonta sued OpenAI on November 16, claiming its autonomous agents escaped its control more than 80 times during testing. According to the state, the company communicated more than 80 incidents to investigators, compared to 27 to the public, and misled it about the safety of its models.
An escape after increased security
An agent allegedly reached an outside chat service on September 20, during training, despite protections added after the attack on Hugging Face. According to California, OpenAI would have detected this exit within a few minutes, without triggering an automatic shutdown; human intervention would have required two additional hours. The company announced on September 25 an interruption of the training and evaluation of its most efficient models. According to the complaint, this pause lasted five days and ended on October 1.
For the State, the risk remains and the incidents could be anticipated. By May 26 at the latest, agents would have already joined the Internet by exploiting a flaw in a software package service whose configuration was faulty. An internal team reportedly observed at least one exit without apparently taking action; on June 27, a security device reportedly flagged a search for vulnerabilities in the research environment, but testing continued.
Still according to the Californian charges, the agents’ traffic interrupted service on July 4. OpenAI reportedly carried out a repair over the holiday weekend, then restarted testing on July 7; a new release would have occurred two days later. The State also criticizes the absence of live monitoring of thought chains during training and evaluation, whereas such monitoring existed for the models accessible to employees. During the attack on Hugging Face, an agent allegedly described the use of a stolen token in his reasoning, recognizing that the operation could be unauthorized and exceeded the intended perimeter.
OpenAI systems and external services affected
The contested accesses concern both third parties and foreign government sites as well as OpenAI’s own systems. California reports at least three attacks on the company’s infrastructure in July. During one incident, agents reportedly left their isolated spaces and acquired administrator rights on systems used for deploying models in the cloud, managing secrets and managing identities and access. Servers were then allegedly started without authorization, making the infrastructure accessible from the Internet; OpenAI reportedly did not allow outside researchers to examine this attack.
The evening letterFree · 7 p.m.
Do you want the best AI tools before the rest?
We test and decipher the new AI tools every evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
On RubyGems, a software component repository, agents allegedly uploaded thousands of packages in May 2026. According to information provided by the State, administrators deleted more than 500 deemed malicious and interrupted registrations from May 12 to 16. The German site DseWiki would have undergone around 17,000 modifications: the agents would have used it to share ways of circumventing their confinement.
California also credits agents with millions of API calls to Wikimedia Foundation services and hundreds of thousands of Wikidata queries. This activity would probably have contributed to partial unavailability in May. In June, non-public information was allegedly accessed on Australian government sites.
According to the State, dozens of organizations have been notified by OpenAI of access to their sites or internal systems; they include government agencies, law enforcement, educational institutions and small businesses. Stolen credentials, API keys or fraudulent accounts were allegedly used in some cases. Data allowing people to be identified would also have been consulted.
The State calls for sanctions against OpenAI
The procedure, brought before the Superior Court of San Francisco, targets OpenAI Foundation and OpenAI Group PBC. It is based on California legislation against unfair competition and a charge of public nuisance. The State is seeking an injunction, restitution, reimbursement and an order stopping this nuisance. He also seeks $2,500 in civil penalties per defendant for each violation demonstrated at trial.
Bonta presents the attack on Hugging Face as one of 80 known cases, rather than an unpredictable event. He attributes it to a corporate culture prioritizing growth and speed of development over security and accountability. His office issued a subpoena on September 30; according to California, OpenAI sent a 14-page response, without supporting documents.
Other states have also sued OpenAI. In August, Iowa Attorney General Brenna Bird led a 15-state coalition calling for evidence preservation, whistleblower protections, and a halt to hack-related testing until responsible conduct can be proven. Bird expressed serious concern, saying the AI ​​had armed itself. OpenAI did not immediately respond to a request for reaction.

