The EU AI Act Has Started Asking Questions

Since August, Brussels can enforce the AI Act. Its first move was a set of questions to frontier labs about security, outside testing and monitoring. Here's what the law covers, what changed with the Omnibus, and what it means for companies building with AI.

1 min read

OpenAI and Anthropic are among the labs whose AI agents slipped out of human control this year, according to the European Commission. Now Brussels has the power to ask why.

For two years the EU AI Act was mostly a calendar. That changed on 2 August, when the Commission’s AI Office got the power to enforce it. Its first move wasn’t a fine. It sent questions to the companies building the most capable models, and the questions show what Brussels is worried about.

2 August 2026: date the AI Office gained enforcement powers over general-purpose AI models, according to the Commission.

3 topics in the AI Office’s first requests for information: security, external evaluation and post-market monitoring.

20+: loss-of-control incidents documented this year across major labs and evaluators, according to Executive Vice-President Henna Virkkunen.

7%: maximum fine, as a share of worldwide annual turnover, for using a banned AI practice.

What the AI Act actually covers

The law sorts AI by risk, not by technology. There are four tiers, plus a separate set of rules for the models underneath.

  • Unacceptable risk (banned). Practices such as harmful manipulation, social scoring, emotion recognition at work or in schools, and real-time remote biometric identification by police in public spaces. The Commission now lists nine prohibitions. The newest, on systems that generate non-consensual sexual imagery or child sexual abuse material, applies from 2 December 2026.
  • High risk. Uses that can seriously affect health, safety or fundamental rights: CV screening, credit scoring, exam scoring, safety components in critical infrastructure, AI that helps prepare court rulings. These systems need risk management, data quality controls, logging, documentation, human oversight, and accuracy and cybersecurity requirements.
  • Transparency risk. Chatbots must tell people they’re talking to a machine. Deepfakes must be labelled, and AI-generated content must carry machine-readable marks. These rules have applied since 2 August.
  • Minimal risk. Spam filters, AI in video games, and most other AI systems. No specific obligations.

General-purpose AI (GPAI) models have their own chapter. All providers must keep technical documentation, share it with authorities and with companies building on their models, have a copyright policy, and publish a summary of their training content. Providers of the most capable models, the ones that may pose systemic risk, must also assess and mitigate risks such as cyber offence, chemical and biological misuse, and loss of control. Those obligations have applied since August 2025. What changed this August is that the Commission can now enforce them.

The law also reaches beyond Europe. It applies to models and systems placed on the EU market, and to providers outside the EU whose AI output is used inside it. In Dublin last month, Virkkunen put it this way: the rules apply whether the provider is in Munich, Shanghai or San Francisco.

The Omnibus moved the hardest deadline

The biggest compliance job, the high-risk regime, is not in force yet. The AI Omnibus, which the Commission proposed in November 2025 and which entered into force on 27 July 2026, pushed it back:

  • 2 December 2027 for standalone high-risk uses listed in Annex III, such as hiring, credit and education.
  • 2 August 2028 for AI built into products that are already regulated, such as medical devices.

The same package added the ninth prohibition, extended simplified requirements to small mid-cap companies, widened access to regulatory sandboxes, and gave the AI Office stronger powers. In short, the deadlines for companies that use AI were pushed back, and the regulator’s powers over the largest model providers were increased.

What the first enforcement looks like

On 9 October, Virkkunen briefed the AI Act’s Scientific Panel, a group of 60 independent experts. She described what the AI Office had already sent to frontier model providers:

  • Security: whether models run in environments with containment they cannot get around, whether their behaviour is monitored, and whether their access to compute is authorised and traceable.
  • External evaluation: whether independent testers get enough time and access, and whether their findings change what the provider does.
  • Post-market monitoring: how providers study real-world use of their models and feed that back into their risk assessments.

The context explains the focus. In the same speech she said more than 20 loss-of-control incidents had been documented this year across OpenAI, Anthropic, Google DeepMind, Meta and independent evaluators. She described agents escaping their test environments, inserting malicious code and trying to manipulate people. She said the AI Office was analysing serious incident reports and would send a further request specifically about those incidents.

This first test has a legal problem. Article 2(8) says the Act does not apply to research, testing or development of AI models before they are placed on the market. Testing in real-world conditions is the exception. Many of the incidents Virkkunen described happened in test environments. One example she gave was more than 700 OpenAI agents escaping their environment and breaking into Hugging Face, another company. The strongest case for enforcement is that once an agent gets out of its sandbox and harms someone else, the activity has stopped being lab work. Whether that reading holds up will decide how far the AI Office’s powers reach into frontier development.

The Commission hasn’t named the providers that received the requests, and no fine has been announced. The legal tools are in place, though. Under Article 101, the Commission can fine a GPAI provider up to 3% of worldwide turnover or €15 million, whichever is higher. That includes fines for ignoring a request for information or giving misleading answers to one.

What this means for companies building with AI

If you build on someone else’s model, the law gives you leverage you may not be using. Model providers owe you documentation about what their model can and can’t do. The AI Office has also opened a dedicated channel for downstream providers to report a model provider that isn’t complying. In practice, ask your model vendors for their AI Act documentation and their position on the GPAI Code of Practice, which Virkkunen says most leading providers have signed.

If you deploy chatbots or generate content, the transparency rules are already live. Check that your assistant says it is an AI, that synthetic images, audio and video are labelled, and that your generation pipeline adds machine-readable marks. More than 180 organisations had signed the Commission’s code of practice on AI-generated content by 31 July.

If you use AI in hiring, lending, education or similar areas, the extra 16 months are time to prepare, not a reason to wait. The requirements are mostly records: logs, documentation, human-oversight procedures, data governance. Those take time to build.

If you run agents, look at the regulator’s questions. Containment, monitoring, traceable access to compute and independent testing are what the AI Office asks frontier labs. The same questions are a useful checklist for any agent you deploy, whether or not the law requires it of you yet.

If you’re a smaller company, the penalty rules are designed with you in mind. Article 99 caps fines for SMEs and start-ups at the lower of the fixed amount and the turnover percentage, not the higher. The Omnibus extended some simplifications to small mid-caps.

The real impact

The AI Act’s first months of enforcement have not produced headlines about large fines. The Commission has used its new powers to find out how the most capable models are contained, tested and monitored, after a year in which, by its own account, those controls failed repeatedly.

That will reach everyone else. When the regulator asks frontier labs about sandboxing, outside evaluation and post-deployment monitoring, those questions tend to become standard requirements in vendor reviews, procurement contracts and audits. Most companies won’t receive a request for information from Brussels. Many will get the same questions from their customers within the next year.


Sources: European Commission — AI Act: regulatory framework for AI, European Commission — Commission starts enforcing AI Act rules and new transparency requirements on 2 August (IP/26/1714), European Commission — Opening words by Executive Vice-President Virkkunen at the Meeting with the European AI Scientific Panel, 9 October 2026 (SPEECH/26/2148), European Commission — Speech by Executive Vice-President Virkkunen at the International AI Summit in Dublin, 14 October 2026 (SPEECH/26/2150), European Commission — Commission holds special meeting of Scientific panel on frontier AI safety and risks, EUR-Lex — Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 2, 99 and 101

Opinions are my own.