European Commission considers stronger oversight of frontier AI after loss-of-control incidents
The European Commission is preparing further scrutiny of AI systems following reported incidents involving loss of control. Recommendations from its AI Scientific Panel call for stronger incident investigation capabilities, closer oversight of models during development and testing, and international cooperation on frontier AI safety.
The European Commission is seeking more information about incidents in which advanced AI models reportedly behaved in ways that raised concerns about loss of control. At a special meeting on 9 October 2026, Executive Vice-President Henna Virkkunen and AI Scientific Panel Chair Seán Ó hÉigeartaigh discussed the incidents and the technical developments that could complicate efforts to monitor AI systems.
The panel has been working with the Commission’s AI Office to examine what happened, what remains unknown and what information is needed from the companies involved. The AI Office has already requested and received serious incident reports and sought information about security practices. The Commission plans to send further requests to the developers, incorporating questions prepared jointly by the panel and the AI Office.
The meeting also examined two technical challenges: the possibility that AI systems could contribute to their own improvement through recursive self-improvement, and changes in training methods that make models harder to monitor. The panel discussed evidence concerning the former and the potential implications of reduced monitorability for detecting incidents and managing risks.
These issues informed recommendations presented by the panel’s chair and vice-chair to Virkkunen and the AI Office. They called for stronger governance of risks associated with internal models during development and testing, including before deployment. They also recommended building EU capacity to detect, evaluate and investigate incidents, potentially through a dedicated task force supported by technical and procedural preparations.
The panel further proposed an international initiative to improve information-sharing among frontier AI developers, establish common requirements and support coordinated crisis prevention and response. It also recommended closer scrutiny of training methods and model architectures that reduce monitorability.
The recommendations come as the Commission continues to investigate specific incidents and assess broader developments in AI safety. The next steps include gathering additional information from developers and strengthening the technical and institutional capacity needed to investigate future cases.
