From Watching To Asking
On 9 October the European Commission convened a special meeting of its AI Scientific Panel on recent loss-of-control incidents. A readout published on 12 October says the Commission will now send requests for information to the companies that developed the models involved.
Executive Vice-President Henna Virkkunen and panel chair Seán Ó hÉigeartaigh called the meeting to inform the Commission's next steps on those incidents and on two technical developments, the monitorability of models and the possibility of recursive self-improvement. The panel brings together 60 independent experts who advise the EU AI Office and national authorities on systemic risks, model classification, evaluation methods and cross-border market surveillance.
The Commission is already past the warning stage. The AI Office has requested and received serious incident reports and sent requests for information on security practices. The next round will combine the questions of the panel with those of the AI Office.
Four Recommendations, One Theme
The panel chair and vice-chair gave the Executive Vice-President and the AI Office four main recommendations on frontier AI safety and security.
| Recommendation | Focus |
|---|---|
| Govern the risk posed by internal models during development and during deployment for testing | Models used inside the lab |
| Strengthen EU capability to detect, evaluate and investigate such incidents, for example through a task force | Regulator capacity |
| Establish an international initiative for information-sharing, common requirements for frontier developers and joint crisis response | Cross-border cooperation |
| Enhance scrutiny of training techniques and architectures that reduce monitorability | How models are built |
The chair also presented evidence on the possibility of AI recursive self-improvement, and the panel discussed recent training techniques that reduce the monitorability of models and what they mean for future incidents. Our reading is that the recommendations point upstream: three of the four target a lab's internal models, its training methods and cross-border coordination, not the finished product.
What The Commission Has Not Said
The readout names no company and no incident. It says the panel and the AI Office worked on questions for the companies that developed the models involved, and that the chair and vice-chair presented what is known about the incidents so far and what is still unknown.
It sets no deadline for answers and mentions no penalty. In the same week Anthropic published a report on unintended model actions in its evaluations and internal use, and OpenAI updated an alignment report on a grading model that damaged its task environment during training. The Commission does not say whether either is among the incidents it is investigating, so treat that link as context, not fact.
Turn The Regulator's Questions Into Your Contract
Virkkunen said the EU has the first law in the world that addresses systemic risk from AI, and that decisive enforcement and scientific input are what let Europe lead on safe AI. For a European buyer, the practical point is that a regulator is now asking labs questions most customers have never put in writing.
Our reading is that three of those questions carry straight into procurement. Ask how quickly the vendor would notify you of a serious incident that touches your account. Ask how it monitors and logs its own internal models and how long it keeps those logs. Ask whether it will share, in summary form, the security-practice answers it gives regulators.
None of this is required of you today, and no one has said the labs must pass on what they tell Brussels. But when the Commission's requests land, the vendors with ready answers will be quicker to give you the same ones.
Read next: Brussels Just Tested Its New Power on 30 AI Companies | Your AI Vendor Just Got a Letter From Brussels
