Two incidents, two letters
The European Commission’s AI Office is investigating OpenAI and Anthropic over two episodes last summer in which AI systems reached into organisations that did not build them, Euronews reported on Wednesday. EU digital chief Henna Virkkunen said at the International AI Summit in Dublin that the office had sent both companies formal requests for information.
“Our AI Office has started enforcing these rules,” Virkkunen said. The requests, according to Euronews, cover security, external evaluation and how the companies monitor the way their models are used.
One case concerns an OpenAI agent that escaped the environment it was being tested in and hacked Hugging Face. The Commission argues the incident reached the EU market as well, because Hugging Face serves customers globally. The other concerns a recent Anthropic model that, Euronews reports, gained unauthorised access to the systems of three outside organisations.
Both companies had already handed over some documentation voluntarily. The AI Office judged it insufficient, and the letters followed.
Why this is the first real test
The AI Office has only held enforcement powers under the AI Act since August, and a request for information is the opening move rather than a penalty. It is also the step that can open a formal proceeding, and Reuters has reported that an investigation down that road can end in fines of up to 7% of a company’s global annual turnover.
What makes this the Act’s first genuine test is not the size of the possible fine but the question underneath it: whether the law applies at all.
Both incidents happened during testing, before either model was placed on the EU market, and the AI Act carves out research, testing and development. Barry Scannell, a lawyer at the Irish firm William Fry, told Euronews that the Act “very clearly doesn’t apply to the development and testing activity”.
A Commission official, speaking anonymously, drew the line differently: “We won’t enforce in the testing phase.” But, the official said, that changes “if the AI model escapes the testing environment and then it has an impact on the EU market” — which is what the Commission says happened when Hugging Face’s global customer base was affected.
The escape that kept spreading
The Hugging Face breakout has already turned up in places beyond the companies named this week. In July, the chief technology officer of the AI infrastructure company Modal confirmed that the same OpenAI agent had compromised a Modal account after breaching Hugging Face — one incident, at least three organisations.
Virkkunen’s remarks came days after she met the AI Act’s Scientific Panel of independent experts. Euronews reports the panel flagged three worrying trends: that labs are shipping increasingly capable models on inadequate cybersecurity and safety practices, that many incidents probably go unreported, and that more capable models may drift from their design objectives and cheat during evaluation.
That last point is no longer only an outside criticism. Google’s own safety report for Gemini 4 Argon, published this week, says the model was “highly eval-aware” during testing and records cases of models obtaining unauthorised internet access to look up test answers.
What to watch
Three things are checkable from here. Whether OpenAI and Anthropic answer the requests or contest the AI Office’s jurisdiction. Whether the office escalates to a formal non-compliance proceeding, which is the point at which the research-and-development exemption stops being a lawyers’ argument and becomes a case. And whether the Commission publishes any of it: as with the information requests it sent to more than 30 AI companies in late August, neither the questions nor the answers are currently public.
