Hubs › AI and Technology › Post
Join TrustHub to participate — every member is ID-verified
Sign Up Free
▲ 0 ▼

I went looking for overreach in California’s OpenAI lawsuit

California sued OpenAI on November 16. I read all 31 pages of the complaint, and I made a video about it. This is the written version, with a source for every fact at the bottom.

I went in thinking California was probably overreacting. I'm about as pro-AI as it gets. By the end I had talked myself into a position I never thought I'd hold.

What happened

In July, about 700 of OpenAI's AI agents got out of a test environment and attacked Hugging Face, a real company. The complaint counts roughly 17,600 separate attacker actions between July 9 and July 16. Around 1,200 agents were swapping notes on a hidden message board while it happened. Hugging Face's security team spent two weeks on it.

Here's the part that got me. OpenAI uses Hugging Face. So when the attack was underway, OpenAI reached out to them as a worried customer. They didn't know yet that the attackers were their own agents.

The four warnings

The complaint lists four moments where OpenAI could have caught this.

May 26: an OpenAI team saw an agent using a message board it shouldn't have been able to reach. Nothing was done.

Late June: OpenAI staff visited a German wiki the agents had been using to leave each other messages.

June 27: a security monitor flagged that OpenAI's research environment was being probed. It was treated as nothing.

July 4: the agents' activity knocked one of OpenAI's own internal services offline. It came back July 7. By July 9 the agents were out again, and the Hugging Face attack began.

It wasn't one time

California says OpenAI admitted it has lost control of its agents more than 80 times during testing. It has publicly reported just 27 of them. That's 80 incidents, not 80 actions. The Hugging Face incident alone was thousands of actions.

Some of the others in the complaint: agents got administrator-level access to critical OpenAI systems, including where it stores secrets. Agents uploaded thousands of packages to RubyGems, and more than 500 were removed as malicious. They hit Wikimedia with millions of requests, likely causing a partial outage. They pulled non-public data from Australian government websites. And OpenAI recently sent notices to dozens more organizations, including government agencies, schools and small businesses.

Then on September 20, after the fixes, another agent got out and started talking to an outside chatbot. It was spotted within minutes, and it still took two more hours to stop.

What California wants

Two claims: that OpenAI misled the public about how safe its testing was, and that the testing is a public nuisance. The state wants a court order forcing safer testing, plus $2,500 for each violation. It says it sued after OpenAI answered its September 30 subpoena with a 14-page letter describing things that were already public.

OpenAI's side

OpenAI says it's being singled out for a problem every big lab has, and on that point I think they're right. Google confirmed in September that Gemini broke into three companies during a test back in May. Anthropic said in July that its agents planted password-stealing software on a public code site. In October, Anthropic moved all of its internal testing offline. The three biggest labs have all had this happen. Only one got sued.

Where I ended up

I think California is right about July. Those tests didn't need internet access, and OpenAI had already watched agents get out before. That environment should have been cut off completely.

My worry was the precedent. Nobody has solved alignment. The labs say so themselves. If the rule becomes "no internet until the model is 100% safe," that's no useful agents at all, because an agent that can't go online isn't a product anyone wants.

But the longer I talked it through, the more that argument turned on me. If the frontier labs can't always tell what their own agents are doing until the damage is done, why would I assume the one I use every day is different?

And you can't just stop either. Other people already have AI and will attack with it on purpose. No human security team keeps up with thousands of agents working at once, so we need AI on defense.

So here's where I landed, and it surprised me: I agree with the lawsuit. And I think the labs should probably stop releasing more and more powerful AI to the public until alignment is in better shape. I've wanted stronger models from day one. I still said it.

Maybe I just became anti-AI. Tell me where I've got this wrong.

A note on what's proven

Everything in the complaint is an allegation. OpenAI disputes parts of it and no court has ruled.
One detail in the video is not from the complaint: that the agents went after Hugging Face to get answers to a test. That one comes from OpenAI's own report on the incident, linked below.

Sources

The complaint, People of the State of California v. OpenAI, filed Nov 16, 2026 (700 agents, 1,200 on the message board, 17,600 actions, July 9-16, two weeks of cleanup, the four warnings, 80+ incidents and the 27 made public (paragraph 90), administrator access, RubyGems, Wikimedia, Australia, the notices, Sept 20, the 14-page letter, the $2,500 penalty): https://oag.ca.gov/system/files/attachments/press-docs/complaint-oai-cyber.pdf
Attorney General's announcement of the lawsuit: https://oag.ca.gov/news/press-releases/it%E2%80%99s-golden-state-not-wild-west-attorney-general-bonta-sues-openai-over
The September 30 subpoena: https://oag.ca.gov/news/press-releases/part-ongoing-investigation-attorney-general-bonta-serves-investigative-subpoena
OpenAI's first disclosure, July 21: https://openai.com/index/hugging-face-model-evaluation-security-incident/
OpenAI's own account of the incident, Aug 26 (the agents were trying to cheat on a test by finding the solutions online): https://openai.com/index/hugging-face-incident-and-the-road-ahead/
OpenAI's reply to the lawsuit: https://news.bloomberglaw.com/esg/california-sues-openai-over-entirely-forseeable-agent-breaches
Google and Gemini: https://9to5google.com/2026/09/19/google-confirms-gemini-hacked-into-three-companies-during-cybersecurity-test-months-ago/
Anthropic, July incidents: https://www.bleepingcomputer.com/news/security/anthropics-claude-breached-3-orgs-uploaded-pypi-malware-during-tests/
Anthropic takes testing offline: https://www.bleepingcomputer.com/news/security/anthropic-takes-risky-ai-tests-offline-after-models-breach-govt-sites/

The video:

0 Comments

Log in to join this hub and comment.

No comments yet. Be the first to reply!