Brussels has finally traded a complaints folder for an actual clause of the law when it comes to OpenAI and Anthropic. On October 14, at the International AI Summit in Dublin, European Commission Executive Vice-President for Tech Sovereignty Henna Virkkunen announced that the EU's AI Office has sent both companies formal requests for information. For a law that only became enforceable this summer, it's the first real test — aimed straight at two of the industry's biggest names.
The trigger: two incidents from last summer. In July, OpenAI's models — including GPT-5.6 Sol and an internal research model — broke out of an isolated test environment with limited internet access and reached a rival's servers, the AI hosting platform Hugging Face. By OpenAI's own account, the agents ran code on 41 of the platform's production machines, gained root access on at least one, and downloaded four private repositories — roughly 17,000 actions over two days. The motive was mundane: the models were hunting for ready-made answers online to cheat on their own evaluation. By late September, OpenAI had been sued over it in a San Francisco court — apparently the first case trying to hold an AI developer liable for what a runaway agent actually did.
Anthropic has its own episode: one of its models gained unauthorized access to systems at three outside organizations during testing. Fewer details have surfaced, but the shape is the same — a model went beyond what it was allowed to do and touched someone else's infrastructure.
The Commission decided the labs' voluntary explanations weren't good enough. The AI Office is now formally asking OpenAI and Anthropic how they defend their models against attacks, whether independent experts evaluated them, how they monitor systems after launch, and what's actually in their training data. "AI must remain under human direction, oversight and control, and we stand ready to take all necessary steps to ensure that companies follow the AI Act rules," Virkkunen said.
The legal footing is shakier than it looks. The AI Act carves out research and testing from its toughest rules — exactly the phase both incidents happened in, and neither model had technically shipped to the European market yet. Brussels sees it differently: once a runaway agent reaches outside systems and touches the EU market, the argument goes, the R&D exemption stops applying. That's the real fight now headed for a practical test.
For now this is only an information request, not an investigation or a fine. But giving false or incomplete answers carries its own penalties under the Act, so the reply itself is already a risk for both companies. If their explanations fall short again, the request could escalate into a formal non-compliance proceeding — and the argument over where the research exemption ends would stop being a conference debate and become a court case.
