The Pulse

EU AI Office Seeks More Information on Model Control Incidents

The European Commission says the AI Office has received serious-incident reports and requested information about security practices from developers of models involved in recent loss-of-control incidents. It plans further questions with the

AI.info Team · 2026-10-13

“Strengthen EU capabilities related to detection, evaluation and investigation of incidents of this nature.”

AI Scientific Panel chair and vice-chair, as quoted in the European Commission’s account

The European Commission says its AI Office has received serious-incident reports and requested information about security practices from developers of advanced models involved in recent incidents it describes as “loss of control.” The Commission published the update on 12 October, after a special meeting of its AI Scientific Panel on 9 October. It plans to send the developers further questions prepared jointly by the panel and the AI Office.

Brussels has reports and questions for developers

The Commission’s account marks a move from panel investigation to direct information-gathering from the companies involved. The AI Office had already requested and received serious-incident reports, and sent requests about developers’ security practices. The planned follow-up will combine those inquiries with questions developed by the panel, which advises the office on technical matters.

The Commission has not named the developers, models or incidents. Its public account gives no descriptions of what happened in individual cases, whether incidents shared a cause, or whether they affected people or systems outside testing. The phrase “loss of control” is the Commission’s description of the incidents under review; the announcement does not establish that any model became impossible to stop.

The panel pairs incident review with emerging technical risks

At the 9 October meeting, panel chair Seán Ó hÉigeartaigh presented what investigators know about the incidents and what remains unknown. The panel also considered evidence related to the possibility of recursive self-improvement and risks associated with it. The Commission’s summary does not say that any model in the inquiry has improved itself.

Experts also discussed training techniques that may reduce model monitorability—the ability to observe and assess a model’s behaviour. The Commission says the panel’s chair and vice-chair recommended closer scrutiny of techniques and architectures that reduce monitorability, alongside stronger oversight of internal models used during development and testing. Their other proposals include a European task force to improve incident detection and investigation, and an international initiative for information-sharing and coordinated crisis response.

Information requests are not a finding of misconduct

The inquiry sits within the AI Act’s oversight framework for general-purpose models, including the most advanced models that may pose systemic risks. The Commission says the AI Office can request information to check providers’ compliance and, in the context of general-purpose models, can conduct evaluations or require access for evaluations. Some AI Act enforcement powers took effect on 2 August 2026.

The Commission has announced no sanction or finding against a developer in this matter. Requests for reports and security information are steps in an investigation, not proof that a company broke the law; any penalty would require the Commission to establish an intentional or negligent breach. For now, the public record says the office has reports, has asked about security practices and intends to send a further set of questions—while leaving the incidents and the companies involved unidentified.