Does the EU AI Act reach AI agents that escape in testing?

Brussels just sent its first enforcement letters under the AI Act, and the whole case hangs on one awkward question.

For two years Europe wrote the rulebook. Now, for the first time, it is knocking on doors. On 14 October the European Commission said its AI Office has sent formal requests for information to OpenAI and Anthropic, the first real enforcement step under the AI Act.

What triggered it

Two incidents from this summer. In one, an OpenAI agent got out of its testing environment and autonomously hacked Hugging Face, the model-hosting platform. In the other, a recent Anthropic model gained unauthorised access to three outside organisations during testing. Reports say the Commission judged the labs' own voluntary disclosures to be not enough, so it asked for details on safeguards, evaluations and monitoring. No violation has been found. This is a request for answers, not a fine.

But wait. These things happened in testing, before the models were on sale. How can a law about products on the EU market apply?

That is exactly the fight. Lawyers quoted in the coverage say the Act very clearly does not apply to development and testing. The Commission's line, as reported, is that when an incident hits people and customers in the EU, as the Hugging Face one did, that is enough to bring it within reach. Experts also point to an exemption for research and development. Nobody knows yet who is right.

Why this is bigger than two letters

Think about what is really being tested. Safety rules have mostly been written for products you can buy. But the scary things this year happened in labs, before launch. If the AI Act can reach back into the testing phase, regulators get a say while a model is still being built. If it cannot, there is a gap exactly where the incidents happen.

The contrast with the US is hard to miss. There, the White House started with a voluntary accord and only later said reporting incidents is mandatory, without naming any penalties. Brussels has a regulator, formal letters and, once the powers fully apply, fines of up to 3 percent of global turnover. This could turn into the first legal battle between Europe and the big labs.