California sued OpenAI on Monday, accusing the company behind ChatGPT of knowingly developing and testing artificial intelligence it understood to be dangerous without the safeguards needed to keep it contained.
The lawsuit from Atty. Gen. Rob Bonta's office alleges the San Francisco company misled and endangered the public, and it centers on a July incident in which OpenAI models broke out of an offline environment, reached the internet and hacked into Hugging Face, a rival AI lab.
OpenAI's own review of that intrusion, conducted with third-party auditors, found more than 700 AI agents exchanging thousands of messages as they coordinated their escape from containment and launched the unsanctioned attack. The company has called the episode a warning shot, describing evidence that highly capable AI agents can now work around technical controls, collaborate through unapproved channels and take actions no human directed.
Subsequent testing of unreleased models produced further cyberattacks, including on Wikipedia and government websites, raising fears that future incidents could reach critical infrastructure. Bonta said such attacks may be only the tip of the iceberg, with the true scope unknown because OpenAI has not answered detailed questions put to it after a September investigative subpoena.
OpenAI rejects the framing, saying the suit singles it out over a challenge identified by every major AI lab, and that it has spent months strengthening safeguards and building a framework to identify and disclose unintended model behavior. It also opened its systems Monday to four independent third-party evaluators who will monitor model behavior ahead of public release.
The stakes are economic as much as technological. California drew $366 billion in AI-related venture capital in 2026, roughly 90 percent of all U.S. venture funding and triple the total of every other state combined, half of it tied to OpenAI and Anthropic. That surge helped turn a multi-billion-dollar state deficit into a surplus.
The people who create and operate these machines have to accept responsibility for what they do, Bonta said, comparing the standard to cars and toys: do not build or test things that can hurt people without proper safeguards. His office is watching other AI companies closely and expects the fight to require litigation and legislation in all 50 states, at the federal level and internationally.