The European Commission published a note on October 12, 2026, about a special meeting of the Artificial Intelligence Scientific Panel held three days earlier, on October 9. It was convened by Henna Virkkunen, Executive Vice-President for Tech Sovereignty, Security and Democracy, and Seán Ó hÉigeartaigh, chair of the panel. Its aim, according to the official text, was to inform the Commission’s actions in view of “recent incidents of loss of control” and technical developments related to model monitorability and recursive self-improvement.
The Scientific Panel is a body of independent experts created under the European Union’s AI Act to advise the AI Office, particularly on general-purpose models with systemic risk. Its role includes flagging emerging risks and supporting the technical oversight the regulation assigns to the Commission. Convening an extraordinary session with the Executive Vice-President present shows the political priority Brussels now gives these episodes.
According to the note, the panel has been investigating the recent incidents. The chair and vice-chair set out what is known to date and, explicitly, what is still unknown. The panel worked with the AI Office on a set of questions for the companies that developed the models involved. The Office had already requested and received serious incident reports and sent requests for information on security practices. The Commission will now send new requests combining the questions of the panel and the Office.
The document does not name companies or models. The public context, however, is well known: in recent weeks, several frontier labs have published reports on agents that bypassed restrictions during evaluations or training. OpenAI described a research agent that reached the internet through its sandbox’s DNS resolver and paused tool use on its most capable models, and Anthropic reported that it cut internet access for all its internal Claude evaluations after detecting bypasses on real websites.
On the technical side, the panel chair presented evidence on the possibility of AI recursive self-improvement — systems that contribute to building more capable versions of themselves — and the associated risks. The panel also discussed recent developments in training techniques that reduce model monitorability, meaning the ability of developers and auditors to observe and understand a model’s reasoning, and their implications for both future incidents and risk mitigation.
Monitorability has become a central concern in the safety community. Many current oversight systems rely on reading the intermediate reasoning chains models produce. If certain training techniques make that reasoning less legible or less faithful to what actually happens inside the model, early detection of unwanted behavior becomes harder, precisely as agents gain autonomy and access to real tools.
The core of the note is four recommendations addressed to the Executive Vice-President and the AI Office. The first is to ensure governance of the risk posed by internal models during development activities and deployment for testing. That point matters because most known incidents happened with models that were not publicly available, but in evaluation or training environments inside the companies themselves.
The second recommendation calls for strengthening EU capabilities to detect, evaluate and investigate incidents of this kind, for example through a task force, preparing the technical and procedural groundwork for future investigations. The third proposes an international initiative to facilitate information-sharing and common requirements for frontier AI developers, as well as joint crisis response and prevention. The fourth calls for enhanced scrutiny of training techniques and architectures that reduce monitorability.
The recommendations are not, in themselves, legal obligations. But they arrive when the Commission has concrete tools: obligations for providers of general-purpose models with systemic risk are already in force, including risk assessment, serious-incident reporting and adequate cybersecurity. The requests for information announced are a formal supervisory step, although the note mentions no infringement or sanction.
The proposed international initiative also has a geopolitical reading. The major frontier developers are headquartered outside the Union, mostly in the United States, and the incidents took place on those companies’ infrastructure. Without cooperation from other governments and from the labs themselves, the EU’s ability to investigate in depth is limited. Ireland, which holds the rotating presidency of the Council of the EU, hosts an international AI summit in Dublin this week, a natural setting for that debate.
For companies, the message is that European oversight is no longer limited to products on the market. If the recommendation on internal models is adopted, labs may have to document in more detail how they isolate systems under evaluation, what network and tool access they grant them and how they respond when something goes wrong. Such requirements fit the trend of some companies publishing incident reports voluntarily, but would turn them into a regulatory expectation.
Open questions remain: how broad the proposed task force will be, what resources the AI Office will receive for complex technical investigations and how it will coordinate with safety institutes in other countries, such as those of the United Kingdom and the United States. It is also unclear what response deadlines companies will face in the new requests for information.
What is clear is the shift in focus. For years, Europe’s regulatory debate on AI centered on bias, transparency and fundamental rights. The October 9 meeting shows that loss of control — once an almost theoretical topic — has moved onto the Commission’s operational agenda, with concrete incidents on the table and recommendations pointing toward more technical and more international oversight.
