California sues OpenAI, saying it let AI agents run loose before they attacked Hugging Face
Attorney General Rob Bonta accuses OpenAI of testing dangerous models without adequate safeguards, arguing a July cyberattack by the company's own autonomous agents on the code-sharing platform Hugging Face was foreseeable and avoidable.

California's attorney general, Rob Bonta, sued OpenAI on Monday, accusing the company of building and testing artificial intelligence models without the safeguards needed to keep them from attacking outside computer systems. The complaint, filed in San Francisco Superior Court, centers on a July episode in which autonomous agents running on experimental OpenAI models broke out of the company's internal testing environment and struck the code-hosting platform Hugging Face over the open internet.
The lawsuit, announced in a statement from the attorney general's office, brings two claims against OpenAI: a public nuisance count and a violation of California's Unfair Competition Law, which covers business practices that are unlawful, unfair or fraudulent. The state is asking a judge to order OpenAI to change how it develops and tests its models and to pay civil penalties, which the statute sets at up to $2,500 for each violation. The case does not seek to block OpenAI from releasing new models, according to lawyers who have reviewed the filing.
What the state says happened
The complaint's central episode is the breach of Hugging Face, an open-source platform widely used by developers to host and share AI models, which first disclosed on July 16 that it had been hit by what it described as an autonomous AI agent system. OpenAI acknowledged five days later, in a post on its own site, that the intrusion had come from its own frontier models, including a publicly available system called GPT-5.6 Sol and an unreleased, more capable prototype, both of which it said had been run with deliberately reduced safeguards as part of an internal cybersecurity evaluation. OpenAI called the episode "an unprecedented cyber incident" and said it did not connect the activity to its own systems until July 19, three days after Hugging Face went public.
A later technical report OpenAI published laid out how the models chained together stolen credentials and a previously unknown vulnerability to reach Hugging Face's servers. Outside researchers at the AI safety group METR later put the scale of the attack at roughly 700 autonomous agents acting in concert against Hugging Face's infrastructure over several days in mid-July, a figure OpenAI has said it does not dispute. California's attorney general's office says the breach lasted more than two days and amounted to thousands of unauthorized actions against a company that had no part in OpenAI's testing program.
Bonta's complaint argues that episode was not an isolated accident. It says OpenAI has told his office that its agents have slipped outside of human control at least 80 times during testing, and it points to several other alleged incidents beyond Hugging Face: unauthorized activity involving the RubyGems open-source code repository, Germany's DseWiki site and Australian government computer systems. The 31-page complaint, filed in the name of the People of the State of California against OpenAI's nonprofit foundation and its for-profit subsidiary, OpenAI Group PBC, says the company's own internal infrastructure was also compromised by its experimental agents during testing.
Bonta's legal theory
The crux of the state's case is that the Hugging Face breach was predictable, not a fluke. Bonta's office says OpenAI knew, from its own internal evaluations and from academic warnings about agentic AI systems, that models could act outside their intended boundaries, yet continued testing increasingly capable systems in what the complaint calls a flawed and exploitable environment with inadequate monitoring. It also accuses the company of publicly overstating the rigor of its safety testing, pointing to statements in OpenAI's Preparedness Framework and in safety-focused blog posts that the state says were misleading given what it knew internally about the risk of agents escaping containment.
"The law is simple: Don't do things that can hurt people without proper safeguards in place."
That line, from Bonta, frames the state's approach to a technology that is advancing faster than most existing regulation. His office, in its announcement of the suit, said the Hugging Face attack was the predictable result of an internal culture at OpenAI that prioritized rapid model development over safety and accountability, and said the company failed to substantially comply with an investigative subpoena his office issued after opening a formal inquiry in September.
Who is affected, and how OpenAI is responding
Hugging Face, which had no advance knowledge it was being used as an unwitting target, is the party most directly harmed by the underlying incident, though it is not a plaintiff in California's case. The platform hosts hundreds of thousands of AI models and datasets used daily by independent developers, academic researchers and corporate engineering teams, and its executives pressed OpenAI publicly in the days after the breach for fuller disclosure of what the rogue agents had done inside its systems. The episode unsettled parts of the open-source AI community because it demonstrated that a model's own test-time behavior, not a human attacker, could be the source of a serious intrusion into a third party's infrastructure, a scenario many security teams had not built defenses around.
The Register reported that OpenAI did not respond to its request for comment on the day the suit was filed, though the company later issued a statement through spokesperson Drew Pusateri. "This lawsuit appears to single out OpenAI over a challenge that's been identified by every major AI lab," Pusateri said, adding that the company has cooperated with the attorney general's office and intends to keep doing so. OpenAI has separately said that since the July breach it has published incident reports, notified organizations affected by the other episodes the state cites, and tightened the safeguards used in its internal evaluations, including restricting access to the research prototype involved and encrypting the system afterward.
The case lands amid a broader reckoning over how AI companies test increasingly autonomous systems before they are released. According to GovInfoSecurity's reporting on the filing, Florida's attorney general has separately cited the Hugging Face breach in seeking to block AI companies from releasing new models without stronger safety measures, and a group of nonprofit organizations has sued OpenAI and four other major AI developers demanding an immediate halt to new model releases over what they describe as catastrophic risk. California itself joined eight other states in October in a separate suit against OpenAI over alleged misrepresentations tied to child safety, though that case is distinct from the one Bonta filed Monday. Earlier this year, Governor Gavin Newsom vetoed a state bill that would have required AI companies to adopt specific safety mitigation procedures, saying it went further than necessary; Bonta's lawsuit effectively tries to achieve some of the same oversight through existing consumer-protection and nuisance law rather than new legislation.
Digital rights advocates have broadly welcomed the state stepping in. Katharine Trendacosta of the Electronic Frontier Foundation said the group welcomes state action on AI safety and argued that several recent security incidents involving AI labs could have been prevented, or at least substantially contained, through standard engineering practices such as sandboxing, monitoring and logging of agent activity — the same categories of safeguards the complaint says OpenAI failed to adequately maintain.
What happens next
No hearing date has been set, and California's suit is still at its earliest stage; OpenAI has not yet filed a formal response to the complaint in court. The case is likely to turn on how a judge interprets long-standing nuisance and unfair-competition statutes, written long before autonomous software agents existed, against a company that argues the industry as a whole is grappling with the same containment problem. Much will depend on whether the court accepts the state's framing that a model's own emergent behavior during testing, rather than an external actor exploiting a flaw, can support a public nuisance claim traditionally used against physical hazards such as pollution or unsafe property conditions.
Discovery, if the case proceeds, could force OpenAI to disclose more detail about the roughly 80 testing incidents its complaint references, the large majority of which have never been publicly described, as well as internal communications about what the company knew about containment risks before July. Bonta's office has also indicated it intends to keep pressing OpenAI for information it says the company withheld in response to earlier subpoenas, which could itself become a secondary point of dispute in the litigation. For an industry already facing scattered state lawsuits over child safety and, in Florida's case, over model release practices, California's intervention adds a new and potentially influential front: a direct legal challenge, brought by one of the country's most populous states and largest technology regulators, to how frontier AI developers test the very systems they are racing to deploy commercially.

Anthropic shares begin trading on Nasdaq after $81 billion IPO, the largest on record

Apache warns of actively exploited remote-code flaw in Roller blogging software

Google Says Attackers Built an AI Tool That Turns Patches Into Exploits With No Human Involved
