EU Scientific Panel on AI meets on loss-of-control incidents, urges international info-sharing

The European Commission reports that on 9 October 2026 Executive Vice-President Henna Virkkunen joined the panel's chair, Seán Ó hÉigeartaigh, in calling a special session of the AI Scientific Panel. Its purpose was to inform Commission action in light of recent loss-of-control incidents and of technical developments concerning how well models can be monitored and the prospect of recursive self-improvement.

According to the Commission, the panel has been examining the incidents and the panel's two leaders set out current knowledge and highlighted open gaps. The panel and the AI Office have prepared a shared set of questions for the companies whose models were involved. The AI Office has already asked for and obtained serious incident reports and has sent requests for information about security practices. The Commission will now send further requests that merge the panel's questions with those of the AI Office.

On technical matters, the chair presented evidence on possible AI recursive self-improvement and the risks attached, and the panel discussed newer training techniques that make models harder to monitor, along with what that means for future incidents.

The chair and vice-chair then gave recommendations on frontier AI safety and security to Virkkunen and the AI Office. They advise governing the risks of internal models during development and testing deployments; building EU capacity to detect, evaluate and investigate such incidents, for instance through a task force that prepares technical and procedural groundwork; setting up an international initiative for information-sharing, shared requirements for frontier developers and joint crisis response and prevention; and paying closer attention to training techniques and architectures that reduce monitorability.