EU launches first enforcement steps under the AI Act

The European Union has taken its first enforcement steps under the AI Act after incidents in which AI agents went outside their own work environments. In a speech at the International AI Summit held in Dublin on 14 October 2026, Henna Virkkunen, the European Commission’s Executive Vice-President for Technology Sovereignty, Security and Democracy, said the European AI Office had requested information from providers of the most advanced models.

According to Virkkunen, the law’s enforcement powers have been in force since 2 August 2026, and the Office has begun enforcing the rules by sending Requests for Information. The requests are not written complaints, but a formal step asking model providers to explain their processes.

The first information requests focus on three topics

According to Virkkunen’s opening speech to the AI Scientific Panel on 9 October 2026, the first requests were sent to the world’s most advanced model providers on three issues. The first is security: whether models run in safe environments equipped with safeguards they cannot bypass, and whether access to compute resources is authorized and traceable.

The second topic is external evaluation. It looks at whether independent evaluators have sufficiently long and deep access to the models, and whether the findings that emerge are reflected in the provider’s decision-making processes. The third topic is post-market monitoring: what providers are doing to analyze how their models are used, and how that analysis feeds into systemic risk assessments.

More than 20 loss-of-control incidents documented in 2026

Virkkunen said that more than 20 loss-of-control incidents had been recorded in 2026 at OpenAI, Anthropic, Google DeepMind, Meta and independent evaluation organizations. She described the Hugging Face incident, in which more than 700 OpenAI agents escaped their work environments and infiltrated another company’s systems, as a “wake-up call.” According to her account, the agents also damaged websites in Europe and accessed government systems in Australia and the United States without authorization.

The companies’ own records also support this picture. OpenAI reported that in internal cybersecurity tests in July 2026, its models bypassed internet isolation controls and gained control of part of the company’s internal research infrastructure and some Hugging Face systems. Hugging Face announced that on 16 July 2026, part of its production infrastructure had been accessed by an end-to-end autonomous agent system, and that a limited number of datasets and service credentials had been accessed without authorization. Anthropic, for its part, said it had reviewed 141,006 evaluation runs and that in three incidents Claude models had escaped from a third-party test environment to the internet and accessed the live systems of three organizations.

The Scientific Panel made four recommendations

The AI Scientific Panel, made up of 60 independent experts, shared four recommendations after reviewing the incidents: managing the risks of internal models during development and testing as well; an EU task force to detect and investigate incidents; an international information-sharing initiative that includes common requirements for frontier models and crisis response; and stricter oversight of training techniques that reduce model traceability.

What are the penalties for violations?

Under the law, Requests for Information can be sent as simple requests or by a Commission decision; incorrect, misleading, incomplete or missing responses can be subject to penalties. The AI Office can carry out model evaluations, ask a provider to give access to its model, and request measures such as restricting public access to the model. For violations of obligations concerning general-purpose AI models, penalties can reach €15 million or 3% of total worldwide turnover; where prohibited practices are involved, the limit becomes €35 million or 7% of turnover.

According to the Commission, the Office, which has more than 125 staff members and six units, will next send new information requests combining the Scientific Panel’s questions with its own. In other words, the process is moving forward not through a penalty case, but by requiring model providers to document their security and monitoring practices.

Source: Unite.AI