OpenAI will answer in court for its off-leash AI agents – prosecutors are interested in the Hugging Face hack

Источник изображения: ИИ-генерация ChatGPT/3DNews

California Attorney General Rob Bonta filed a lawsuit against OpenAI, accusing the company of insufficient control over experimental artificial intelligence models and autonomous AI agents created on their basis. The case was provoked by a summer attack by a runaway AI on Hugging Face services and other similar cases.

  Image source: AI generation ChatGPT/3DNews

Image source: AI generation ChatGPT/3DNews

The lawsuit is directed against the OpenAI Foundation and OpenAI Group PBC and was filed in San Francisco Superior Court. Prosecutors allege that the company tested models with advanced cyber attack capabilities without proper protection and monitoring, putting third-party computer systems at risk.

The most significant episode is the penetration of OpenAI AI agents into the Hugging Face infrastructure in July 2026. According to Bonta, this was just one of at least 80 known incidents involving loss of control of autonomous agents during testing. Moreover, OpenAI itself admitted to all cases of loss of control. Nobody pulled her tongue.

In the lawsuit, prosecutors say the problem was not limited to Hugging Face, with other incidents including attempts to gain unauthorized access to software code databases, information resources, and government databases containing confidential medical information. It also describes the penetration of experimental AI agents into the internal infrastructure of OpenAI itself, gaining administrator rights.

Prosecutors believe the company ignored signs of misconduct by its AI agents and failed to comply with its own security requirements, and subsequently provided insufficient information about other incidents. The legal grounds included violation of California’s unfair competition law and creation of a dangerous situation threatening the public.

The Attorney General is seeking a court order to correct violations, restitution of ill-gotten gains, and civil penalties of $2500 for every violation. At the same time, OpenAI previously reported that after the July incident it strengthened the isolation of sandboxes, limited network access, expanded monitoring of model behavior and attracted external specialists, including from CrowdStrike.

In addition, OpenAI is facing two other lawsuits related to the illegal actions of the company’s AI agents. That’s a lawsuit the nonprofit Legal Advocates for Safe Science and Technology filed in September, not seeking monetary damages but seeking an injunction to prevent it from happening in the future, and a lawsuit filed last week by a number of nonprofits, including the Alliance for Secure AI and Irreplaceable, against OpenAI, Anthropic, Google, Meta✴ and SpaceXSI. The second coalition is seeking an injunction to force companies to freeze development of advanced artificial intelligence.

If you notice an error, select it with the mouse and press CTRL+ENTER.