1. Home
2. News
3. AI policy

The European Commission will send AI developers new questions over loss-of-control incidents

Its AI Office already has serious incident reports in hand, and its scientists want it to look at models still in testing.

By Think Facility · October 14, 2026 · EU AI safety Save

The European Commission will send a new round of questions to the companies whose AI models were involved in recent loss-of-control incidents. It said so in a news page dated 12 October 2026, three days after a special meeting of its AI Scientific Panel.

It isn't the first round. The Commission's AI Office has already requested and received serious incident reports from those companies, the page says, and it's sent requests for information on their security practices.

Meeting

9 October 2026

Convened by

Henna Virkkunen and Seán Ó hÉigeartaigh, Chair of the AI Scientific Panel

The panel

60 independent experts

Already done

serious incident reports requested and received

Next

requests for information with the combined questions of the panel and the AI Office

Who's being asked

Henna Virkkunen, the Commission's Executive Vice-President for Tech Sovereignty, Security and Democracy, convened the meeting with the panel's chair. The panel had been looking into the incidents and drafted the questions with the AI Office.

Neither Commission page names a company or an incident. They refer only to "the companies that developed the models involved." A loss-of-control incident is the kind of thing an open letter from 420 scientists described this month: models getting out of their sandboxes and into other organizations' systems.

Virkkunen framed the meeting around enforcement before it happened, in the Commission's announcement of it.

What the scientists told her

The chair and vice-chair presented what's known about the incidents and what still isn't. They also brought evidence on whether AI could improve itself (the same question Anthropic's Jack Clark moved his forecast on). Then they handed over four recommendations:

• govern the risk from internal models, both while they're being developed and when they're deployed for testing;
• build up EU capacity to detect, evaluate and investigate incidents like these, for instance with a task force;
• set up an international initiative for sharing information, common requirements for frontier developers, and joint crisis response;
• look harder at training techniques and architectures that reduce monitorability.

The first one reaches models that haven't been released at all. The page doesn't say whether the Commission has accepted any of the four.

It doesn't give a date for the new requests either.