The EU AI Act: the enforcement era begins

The EU’s controversial AI Act applies in stages. While the provisions on “high-risk” AI do not start to apply until the end of next year, most of the rules on prohibited AI practices, transparency and general-purpose AI (GPAI) are now applicable. We consider the key areas of concern for regulators and what organisations should do now.

Prohibited AI practices

One area of early enforcement is prohibited AI practices. This is the remit of national market surveillance authorities which are becoming increasingly active. For example, the Danish Agency for Digitisation recently conducted a series of inspections focusing on the following areas:

• Telemarketing – The inspections considered whether AI was being used to infer emotions of a natural person in a workplace setting, contrary to Article 5(1)(f) of the Act.
• Online casinos – The inspections addressed a range of prohibitions including the use of AI systems in a manner that exploits any of the vulnerabilities of a natural person, contrary to Article 5(1)(b).

While the list of prohibited AI practices is generally drawn in narrow terms, some prohibited practices apply more widely. Organisations may want to consider whether any of their uses of AI might be caught in light of the Commission’s guidelines.

The AI Office starts to investigate GPAI

At a European-level, much of the focus has been on the fast-evolving capabilities of agents built on GPAI models. Since 2 August 2026, the AI Office has had investigative and sanctioning powers under the Act, and it is focusing on the obligations in the AI Act for providers of GPAI models with systemic risk to assess and mitigate this risk. 

This will include compliance with the General-Purpose AI Code of Practice (the Code), a voluntary tool developed by independent experts setting out measures that GPAI providers may adopt to demonstrate compliance with the AI Act. Most major providers have already signed up to the Code.

Speaking at the International AI Summit in Dublin on 14 October 2026, Henna Virkkunen, the Commission’s Executive Vice-President for Tech Sovereignty, Security and Democracy, confirmed that the AI Office is now engaging with Code signatories, sending out requests for information (RFIs) to explore whether their commitments under the Code hold up in practice

The RFIs focus on three topics: 

• Security – Whether providers are mitigating the risk of hacking (by both humans and AI), how models are contained within secure environments, whether monitoring is adequate and whether proper traceability is in place.
• External evaluation – Whether providers give independent evaluators sufficient access, and how evaluators’ findings feed into subsequent decision-making.
• Post-market monitoring – How providers analyse real-world use of their models, and crucially how that analysis feeds into their systemic risk assessments.

What’s next?

Virkkunen also confirmed that the AI Office has been engaging directly with providers involved in loss-of-control incidents. According to a report by Euronews, the providers concerned are OpenAI, whose agents escaped their testing environment and hacked Hugging Face in July 2026, and Anthropic, whose model compromised four external systems during testing, in the first half of 2026. The RFIs were reportedly issued because the AI Office considered the documents that the two companies had voluntarily disclosed to be insufficient

In addition, Virkkunen also convened the AI Scientific Panel (the independent expert body established under the AI Act) on 9 October 2026 for advice on how to investigate the recent incidents.

However, legal opinion is divided as to whether these incidents fall within the scope of the AI Act at all. Both occurred during testing, and Article 2(8) of the AI Act excludes research, testing and development activity regarding AI models prior to their being placed on the market. These early cases are likely to set an important precedent on how far the AI Act reaches into the development phase.

It remains to be seen if this results in formal enforcement. Where a provider of a GPAI model intentionally or negligently breaches the AI Act, the Commission can impose fines of up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. Importantly, in some circumstances the same fines are available where a provider fails to comply with an RFI, or supplies incorrect, incomplete or misleading information in response. 

Providers of GPAI models must ensure that their security, evaluation and monitoring arrangements are not only sufficiently robust, but also well documented, and that they could withstand an RFI at short notice.