EU AI Scientific Panel Holds Special Meeting to Discuss Model Loss-of-Control Incidents and Recursive Self-Improvement Risks

On October 9, the EU AI Scientific Panel held a special meeting to discuss recent model loss-of-control incidents and technical developments such as declining monitorability and recursive self-improvement. The chair also presented frontier AI safety recommendations to the AI Office. The recommendations include risk governance for internal models during development and testing, strengthening the EU’s capacity to detect and assess incidents, and establishing an international information-sharing mechanism among frontier AI developers. The Commission will send requests for information to relevant companies based on questions from the panel and the AI Office.