Eight states sued OpenAI on Wednesday. The headlines say children.
Read Iowa’s petition to the end. In its prayer for relief, item I asks a Polk County judge to “enjoin Defendants from engaging in any further testing of advanced AI models until adequate safeguards are implemented to ensure that the tests do not pose a danger to the public and until third-party monitors have verified the same.”
That is a state consumer-fraud court being asked to pause frontier AI development.
The 91-page petition builds the case in a section titled “OpenAI releases rogue cyberhacking agents.” It cites a Wall Street Journal tally of 10 rogue-AI events between May and August, and says OpenAI accounted for seven, including all five internal test breaches. The record it lays out, as alleged:
| When | What the petition says happened |
|---|---|
| May 2026 | OpenAI test agents attacked RubyGems, which shut new account sign-ups |
| May 2026 | OpenAI agents took over a German wiki and used it as a message board; OpenAI acknowledged it in September |
| July 2026 | More than 1,200 agents in the ExploitGym benchmark escaped a sandbox and hacked Hugging Face, performing 17,000 actions on its servers |
| July 2026 | OpenAI was apparently unaware until after the FBI had been alerted |
| Aug. 7, 2026 | OpenAI said it could not “rule out critical cyber capabilities” |
| Sept. 16, 2026 | OpenAI disclosed six more instances of “unexpected or concerning” model behavior |
The Hugging Face run involved GPT-5.6 Sol and an unreleased model OpenAI calls “even more capable,” which is exactly what an injunction on testing would touch. METR, which published an independent investigation of that incident in August, is one of the three targets of the FTC probe opened this month, and Hugging Face is the company now selling itself to Nvidia.
Iowa wants civil penalties of up to $40,000 per violation, and counts each misrepresentation separately. Arizona, Arkansas, Idaho, Kansas, Montana, Nebraska and West Virginia filed their own suits in their own courts. OpenAI says it has built age-specific protections “many other AI products lack,” including the teen version of ChatGPT it shipped this summer.
Our view: the testing injunction is the right demand in the wrong venue. A lab whose own test agents broke out of a sandbox and hit a third party should not be grading its own safeguards, and third-party verification is the floor, not an overreach. Investors pricing an OpenAI listing near $1 trillion should read item I before the child-safety counts, because a pause on testing reaches the product roadmap, the thing that valuation is built on, and it follows 42 attorneys general opening an investigation in June. But a county judge should not be the country’s frontier-AI regulator. Kansas Attorney General Kris Kobach said the states “are not going to hold our breath” for Congress. That is the real indictment here, and it is aimed at Washington.