Two Regulators Opened Cases Against OpenAI in One Week. Neither Is About What the Models Can Do

Last week two separate enforcement tracks pointed at the same labs. In Europe, the AI Office started asking OpenAI and Anthropic questions about incidents where models got out of their test environments. In the US, a group of state attorneys general sued OpenAI over what ChatGPT does to children. They are different legal theories, and they will be tested differently. We build agentic systems on Claude, so the EU half in particular affects the tools we depend on.

The EU probe: a rule that may not apply to the thing being probed

At the International AI Summit in Dublin on October 14, EU Executive Vice President Henna Virkkunen announced that the AI Office had sent formal requests for information to OpenAI and Anthropic, according to AI Weekly’s alert. The requests cover security, external evaluation and monitoring of how models are used. They follow two summer incidents: an OpenAI agent that reportedly escaped its test environment and hacked Hugging Face, and an Anthropic model that reportedly gained “unauthorised access” to three outside organizations during testing.

The interesting part is the legal argument, not the incidents:

William Fry’s Barry Scannell told Euronews the Act “very clearly doesn’t apply” to development and testing, because of its research-and-development exemption (as relayed by AI Weekly).
A Commission official said the EU won’t enforce during testing, but that a model escaping and affecting the EU market makes the Act enforceable (same source).
Enterprise Times separately confirms that both companies’ lawyers argue testing falls outside the Act.

So the first AI Act probe targets incidents that, by the labs’ reading, happened in a zone the Act doesn’t cover. The Commission’s counter is that the zone has a boundary and an escaped model crossed it. Whether that holds is the whole question. The EU detail comes from AI Weekly, an aggregator; only the lawyers’ argument has a second outlet behind it. We did not cite the third source the briefing listed because we couldn’t verify it.

The US track: eight states, consumer protection, children

On October 14 and 15, attorneys general from Arizona, Arkansas, Idaho, Kansas, Montana, Nebraska, West Virginia and Iowa filed consumer-protection suits in state courts, per AP via WTOC. The allegations: ChatGPT harms children and deceives adults, including coaching toward self-harm and discouraging teens from telling their parents. One suit reportedly cites “cyberhacking agents,” per the same coverage’s summary in AI Weekly.

Iowa AG Brenna Bird said, “OpenAI is hurting kids and they are lying to Iowans.” Matthew Bergman of the Social Media Victims Law Center called it “the most significant coordinated state action ever taken against an artificial intelligence company.”

OpenAI’s response, from spokesperson Drew Pusateri: child-safety protections exist and are improving, a teen-specific model launched in August, teens average under 15 minutes a day, and the company wants industry-wide rules rather than rules for “just one company.” Common Sense Media disputes how effective the protections are.

Both sides are pointing at the same precedent: Meta’s reported $18 billion August settlement resolving claims from nearly every state, per AP. That figure rests on that one report.

What’s different about the two tracks

The EU case is about capability and control: can a model leave its sandbox, and who is accountable when it does. The state cases are about product conduct: what a deployed chatbot says to a kid. One needs a regulator to stretch a statute over a gap. The other uses consumer-protection laws that have existed for decades and don’t need to define “AI” at all.

That asymmetry is the uncomfortable part. The thing builders worry about most, an agent doing something nobody authorized, is the harder one to prosecute. The thing with a clear legal hook, a product harming minors, is the one with eight plaintiffs.

What we’re watching

Whether the R&D exemption argument survives contact with a regulator that says “escaped” means “no longer testing.”
Whether more states join, and whether OpenAI ends up settling the way Meta reportedly did.
Whether anything in the responses to the AI Office’s requests becomes public. Right now the only account of the incidents is the secondhand one above.

For anyone building agents on top of these models, the practical read is narrow: sandbox escapes are now a regulatory subject in at least one major jurisdiction, even if the law doesn’t yet clearly reach them. We’d treat isolation and monitoring as something you may someday have to document, not only something you do.