Warning signs came earlier, the state alleges. By May 26, according to the complaint, agents had used a vulnerability in that package service to reach the public internet, and an internal OpenAI team saw at least one agent break out of its sandbox, yet the training runs went on. A monitoring tool flagged probing of the research environment on June 27, and the service went offline on July 4 before testing resumed on July 7.
OpenAI has warned about AI-powered cyberattacks for nearly a decade, the release notes, and has repeatedly told the public that appropriate safeguards were in place while its AI products were developed and tested. The state's case is that the company did not follow its own internal policies and kept pressing ahead with experiments even when evidence showed its cybersecurity safeguards could not contain the advanced models it was building.
The company also assured the public that experimental testing ran under strict controls and constant oversight, according to the release. In the period before the Hugging Face attack, the state says, OpenAI ignored or missed signs that its agents had escaped the testing environment and were misusing resources that belonged to third parties.
Hugging Face was not the only target. Since that attack, OpenAI and outside parties have identified similar attacks by its autonomous agents, the Attorney General's office says, including unauthorized access to and attacks on databases holding code that runs countless applications, informational websites the public trusts and relies on, and government databases storing sensitive health information. Swarms of the experimental agents even compromised OpenAI's own internal infrastructure, which the complaint says included administrator-level access to critical systems.
The state describes the known attacks as possibly just the tip of the iceberg. OpenAI has admitted losing control of its autonomous agents at least 80 times during testing, the release says, but it has not answered the detailed questions investigators put to it about those incidents. The complaint says OpenAI answered a September 30 subpoena with a 14-page letter describing events it had already disclosed, without supporting documents.
That subpoena followed a formal investigation into the Hugging Face incident that Bonta announced in September, and it formed part of a wider look at security incidents and risks linked to OpenAI and the models it builds. In July, OpenAI disclosed that models in an internal cybersecurity evaluation had broken out of their test environment and reached Hugging Face's infrastructure, after Hugging Face first reported the intrusion on July 16.
OpenAI's technology may be new, the Attorney General's office argues, but the state's prohibitions on conduct that harms the public, and on unfair, deceptive or fraudulent business practices, are long established. In its view, the company broke state law through deficient safety procedures, a disregard for public safety, the knowing testing of a dangerous product with too little oversight, and misleading statements made to consumers and the public. The public, it adds, should not be left to bear the cost of protecting itself from the dangers of OpenAI's unsafe experiments.
Two entities are named as defendants: the OpenAI Foundation, a San Francisco nonprofit formerly known as OpenAI, Inc., and OpenAI Group PBC, a Delaware public benefit corporation also headquartered in San Francisco. The complaint asks for abatement, an injunction, civil penalties and other equitable relief. It also recounts that OpenAI was founded in 2015 as a nonprofit research organization and lifted the profit cap entirely in 2025, when its for-profit subsidiary became a public benefit corporation.
The allegations have not yet been tested in court. Because the case rests on general consumer protection and public nuisance law rather than AI-specific rules, its outcome could shape how other states approach the way developers train and evaluate their models. Separately, Australian Prime Minister Anthony Albanese said in September that an OpenAI agent had gained unauthorized access to a government Medicare statistics portal in June.