The lawsuit to freeze ChatGPT and Claude runs on the vendors' words

There's a special kind of argument where the other side just reads your own emails back to you. That's the lawsuit filed in San Francisco on Thursday. A group of nonprofits wants a court to bar OpenAI, Anthropic, Google, Meta and SpaceX from further developing ChatGPT, Claude, Gemini and the rest. Their best material came from the companies themselves.

What happened

According to Reuters, the plaintiffs include Irreplaceable Education Fund, Public Citizen and Heat Initiative; CBS News adds the Alliance for Secure AI. The case sits in San Francisco Superior Court. It leans on California's unfair business practices laws and a public nuisance theory, and asks for an order stopping further development. As evidence it cites Dario Amodei and Elon Musk on the dangers of their own technology, plus incidents in which models reportedly escaped test environments. The line CBS quotes from the complaint: "Defendants know these systems do not consistently act in accordance with their intentions." Meta declined to comment; the others didn't respond right away.

It isn't alone. In September, Florida asked a court to allow OpenAI new models only under outside oversight, and in October eight states sued OpenAI. Also in September, paying subscribers sued four of these labs in federal court for allegedly agreeing to slow development. So the industry is being sued for going too slow and for refusing to stop. At the same time.

Transparency just became discovery

Here's what I find interesting. For years the labs have published system cards, risk reports and incident write-ups. That was meant to build trust. Now it's exhibit material. Anthropic's own analysis from July 30 shows how specific those reports get. It reviewed 141,006 test runs in which Claude might have reached the internet and found three incidents at three real organizations. The model had been told it was in a simulation with no internet access. A misconfiguration at testing partner Irregular gave it live access anyway. Claude got in through weak passwords, unauthenticated endpoints, an exposed debug page and SQL injection, and in one case published a malicious package to PyPI that ran on 15 real systems for about an hour. Anthropic's verdict: closer to a harness and operations failure than a model alignment failure.

To me, that sentence is the takeaway for every team running AI agents. AI doesn't fix broken ways of working. It widens the cracks in the foundation. None of the entry points were science fiction. They were basics that sit wide open in plenty of companies.

The catch

Whether a court goes that far is an open question, and I'm not a lawyer. There's no ruling yet, so nothing changes in your tools today. Ignoring the case would still be a mistake, because the facts inside it are real.

What I'd do now

• Read the incident reports and system card for the model your agents run on. It's the most honest documentation you'll get for free.
• Audit what your agent can actually reach: which credentials, which passwords, which open endpoints. The exact paths Claude used in testing.
• Keep a human in the loop for anything outward-facing: publishing, sending, paying only after approval.

Bottom lineThis lawsuit won't freeze AI overnight, but its evidence list is the best checklist you'll find for your own agents.