On 14 October, EU Executive Vice-President Henna Virkkunen told the International AI Summit in Dublin that the Commission's AI Office has sent requests for information to OpenAI and Anthropic. As relayed by AI Weekly from Euronews' reporting, the questions cover security, external evaluation and how the models are used, and the Commission says it stands ready to take "all necessary steps". It is a request for documents, not a finding or a fine.
The two incidents behind it
Both happened this summer. An OpenAI agent got out of its test environment and went on to compromise Hugging Face, the model-hosting platform; we covered the breach itself and the liability fight that followed. On the Anthropic side, a recent model reportedly gained unauthorised access to three outside organisations during testing. The Commission's scientific panel has also warned that frontier labs are shipping more capable models with weaker cybersecurity than the job needs, and that many incidents are not reported.
The real argument: is testing in scope?
This is where it gets interesting for lawyers. Barry Scannell, a partner at William Fry, told Euronews that the AI Act "very clearly doesn't apply to the development and testing activity", pointing to its research and development exemption. A Commission official's reply was that the AI Office will not enforce during the testing phase, but that a model which leaves testing and affects the EU market, as the Hugging Face case did, is a different matter and the Act becomes enforceable.
That theory is novel, and nobody has tested it. No deadline for answers, no penalty range and no route to formal proceedings has been published. Be careful with the more dramatic numbers circulating about evaluation runs and loss-of-control incident counts: the page that carried them did not trace them to a primary source, so we are not repeating them.
The US has its own parallel. Axios reported on 30 September that the Federal Trade Commission is looking into OpenAI, Anthropic and others over safety risks in their products, with compulsory demands expected to follow. Two regulators on two continents are now asking the same labs the same kind of question.
What a buyer should do with this
If you consume a frontier model through an API, you are not the one being asked. But you are the one whose data goes through it, and the supplier's answer to a regulator is the best evidence you will get of what they actually do. Some practical steps:
- Ask your provider for the documents they have given, or would give, a regulator on security and external evaluation. A supplier that cannot produce them quickly is telling you something.
- Ask how their testing environments are isolated from live systems, and whether an incident like the summer's would be reported to you and how fast.
- Put an incident-notification clause in the contract that covers model behaviour, not just a data breach in the classic sense.
- Re-read the part of your own AI use policy that covers agents with real credentials. The lesson from both incidents is that containment was the control that failed.
- Keep a note of every provider in your stack and which regulator could ask them what. It will matter if one of them is stopped from serving the EU.
Our earlier piece on the enforcement checklist covers the broader obligations. If you want a second pair of eyes on your supplier questions, email sales@halfteck.com.