EU AI safety unit gains powers to fine and ban frontier models from Sunday

From Sunday, a 37-person unit of civil servants in a Brussels office building becomes the world's first standing regulator with the power to fine, restrict or ban the most advanced artificial intelligence models. The European Commission's AI safety unit, known internally as A3, acquires its full arsenal of enforcement powers under the AI Act on 2 August 2026, nearly two years after the legislation entered into force.

The unit's mandate covers the providers of frontier AI models, among them OpenAI, Anthropic, Google and Chinese competitors such as Moonshot and DeepSeek. If the Commission concludes that a company has failed to assess or mitigate systemic risks, it can demand mitigation measures, order the restriction or withdrawal of a model from the market, and impose fines of up to €15 million or 3% of global annual turnover. The rules apply regardless of where the company is headquartered, so long as its models are placed on the EU market.

Four systemic risks define the unit's remit

The AI Act, which became law in August 2024, identifies four categories of systemic risk that frontier model providers must address: the use of AI to enable chemical, biological or nuclear attacks; loss of control over an AI system; AI-driven cyber attacks; and large-scale manipulation of populations. A year ago, the Commission singled out these risks as the basis for the obligations now coming into force. The legislation requires providers to "assess and mitigate possible systemic risks" before and after putting models on the market.

The regulatory trigger is a "serious and substantiated concern of a systemic risk." At that point the Commission can request information, demand access to the model for evaluation, impose corrective measures, and ultimately restrict or recall the model. The fine regime applies when companies refuse information requests, deny model access, or fail to implement mandated measures. The turnover-based penalty, 3% of global annual revenue, is designed to be meaningful even for the largest US tech groups.

A unit built in silence, led by a career bureaucrat

Matthieu Delescluse took charge of unit A3 in January 2026, ending an 18-month period during which the AI Office's overall head, Lucilla Sioli, ran the safety team on an interim basis. Delescluse is a long-serving Commission official whose background spans telecoms policy, roaming regulation and the allocation of EU technology and innovation funding. He holds a PhD in neuroscience from Pierre and Marie Curie University in Paris and studied at the College of Europe in Natolin, but he has no prior specialisation in AI safety.

At a London event organised by the research non-profit FAR.ai earlier this year, Delescluse outlined his approach: "We have the means ... to push the companies to use the state-of-the-art in evaluation and assessing the risk of their models." He confirmed that his unit is already "in dialogue with most of the providers" and warned that formal information requests, model access demands, corrective measures and fines are available "if needed and where appropriate." The Commission declined multiple interview requests for Delescluse.

Inside the team: former Big Tech staff and recent departures

The Commission does not disclose the names of lower-level officials, but several have surfaced. Simon Möller, listed on LinkedIn as "AI safety workstream lead," worked for Google in Zürich before joining the unit, a hire that gives the regulator direct insight into how frontier labs operate. Wout Schellaert conducted "general safety-focused evaluations" as a consultant for OpenAI in 2023 before moving to the Commission in early 2025; he has since left. The turnover underscores a broader concern: whether a public-sector unit can retain the technical talent it needs when private-sector salaries are multiples higher.

Jimmy Farrell, EU AI policy lead at the think tank Pour Demain, argued that funding alone will not solve the recruitment problem. "Providing sufficient resources is one thing," he said, "but it's not everything." The Office must also demonstrate "administrative innovation", competitive salaries, secure contracts and direct communication channels, if it hopes to attract and keep researchers capable of evaluating the most advanced models.

The code of practice that Meta refused to sign

Last summer, unit A3 issued a voluntary code of practice intended to help the largest AI companies comply with their obligations under the Act. The exercise did not go entirely to plan: Meta declined to sign. The company's refusal signalled early friction between the industry and the regulator over what constitutes adequate risk mitigation. Since then, the unit has been preparing for the formal powers that take effect this weekend, while the Commission confirmed on Friday that it is recruiting contract agents for two AI Office units, with headcount expected to rise significantly through 2027.

Industry response: cooperation with caveats

Tom Duff Gordon, OpenAI's vice president for policy in Europe, said his company "has worked together with the EU and other industry players in shaping how the AI Act is implemented" and will "continue working together." He added: "We believe responsible AI can help drive Europe's competitiveness and prosperity." Anthropic and Google have not issued public statements on the unit's new powers, but both have participated in the code-of-practice discussions. The Chinese firms Moonshot and DeepSeek, also in the unit's crosshairs, have not commented.

Why the timing matters: pre-deployment scrutiny

The unit's powers arrive at a critical juncture for frontier AI development. The most consequential regulatory intervention will occur just before a new model is launched: the Commission can demand access to the model, evaluate its risk profile, and require changes before it reaches the market. This pre-deployment scrutiny is a deliberate design choice, the Act aims to catch systemic risks before they scale, rather than reacting after harm occurs. Whether a 37-person team can meaningfully evaluate the most complex models ever built, on timelines that match industry release cycles, remains an open question.

A regulatory experiment the world is watching

No other jurisdiction has yet built a comparable structure. The United States relies on voluntary commitments and agency guidance; the United Kingdom's AI Safety Institute has technical expertise but no statutory powers; China's algorithm registry focuses on content control rather than systemic risk evaluation. The EU's model, a dedicated civil service unit with investigative teeth, backed by a comprehensive risk taxonomy, is being studied in Washington, London, Tokyo and Beijing. Its success or failure will shape the next generation of AI governance globally.

The unit's ability to hire and retain the researchers it needs may prove the decisive factor. Delescluse's team includes people who have worked inside the labs they now regulate, a necessary advantage, but one that depends on those individuals staying in post. The Commission's Friday announcement of expanded hiring for 2027 suggests it recognises the gap. Whether "administrative innovation" can close it before the next wave of model releases is the question that will define the AI Act's credibility.