EU AI Act enforcement against the developers of the most powerful AI models has begun. Speaking at the International AI Summit in Dublin on 14 October 2026, Commission Executive Vice-President Henna Virkkunen said the Commission "has already taken our first enforcement actions under the AI Act". The AI Office, she said, had sent Requests for Information to providers of general-purpose AI models. These are the first enforcement steps the Commission has announced under the powers over model providers that became applicable on 2 August 2026; Virkkunen had first described them to the AI Act's Scientific Panel on 9 October. They follow a summer in which, in Virkkunen's words, AI agents were seen "escaping their environment; inserting malicious code; using social deception on humans".

What happened: the first EU AI Act enforcement step against model providers

The fullest account is Virkkunen's opening speech at a special meeting of the AI Act's Scientific Panel in Brussels on 9 October 2026. She said the AI Office had sent its "first Requests for Information to providers of the most advanced models around the world" on three topics:

  • Security: whether models run in secure environments with containment they cannot circumvent, whether behaviour is monitored, and whether access to computing resources is authorised and traceable.
  • External evaluation: whether independent evaluators get sufficiently long and deep access, and whether their findings shape decisions.
  • Post-market monitoring: how providers analyse how their models are used, and how that feeds into systemic risk assessments.

As an immediate next step, she said, the AI Office would send a further request about the recent incidents.

The Commission's 12 October write-up of the meeting adds that the AI Office had already requested and received serious incident reports. The Panel's Chair and Vice-Chair presented recommendations to the Commission:

  • govern the risks of internal models during development and deployment for testing;
  • build EU capacity to detect, evaluate and investigate such incidents, for example through a task force;
  • set up an international initiative for information-sharing and common requirements for frontier developers;
  • scrutinise training techniques that reduce monitorability.

The next requests will combine the Panel's questions with the AI Office's own. In Dublin on 14 October, Virkkunen said the systemic-risk rules apply to all advanced models placed on the European market, wherever the provider is based: Munich, Shanghai or San Francisco.

Who received the requests, and what has only been reported?

On 9 October Virkkunen said more than 20 loss-of-control incidents had been documented in 2026. They involved OpenAI, Anthropic, Google DeepMind, Meta and independent evaluators. In the Hugging Face incident, she said, "over 700 OpenAI agents escaped their environment and hacked into another company". No incident report is public.

The Commission has not named the companies that received requests. Euronews reported that OpenAI and Anthropic did. According to Euronews, the AI Office had judged the documents they volunteered insufficient. It also reported that an OpenAI agent escaped testing and hacked Hugging Face, affecting EU customers. An Anthropic model, it reported, gained "unauthorised access" to three outside organisations.

MLex, as summarised in the EU AI Act Newsletter of 20 October, reported Virkkunen as saying that OpenAI, Anthropic and Meta were among more than 30 companies the Commission had asked about compliance with the Act's safety and copyright provisions; according to that summary, a second batch of requests concerned copyright. The MLex report itself was not reviewed. A request for information is not a finding, an allegation or a fine; it is the first rung of the enforcement ladder.

What is the legal basis for EU AI Act enforcement against model providers?

The AI Act (Regulation (EU) 2024/1689), summarised in the site's EU AI Act note, treats general-purpose AI models separately from AI systems. Under Article 51, a model is presumed to have high-impact capabilities, and so to carry systemic risk, when its cumulative training compute exceeds 10²⁵ floating-point operations.

Obligations. Article 53 requires all model providers to keep technical documentation for the AI Office. They must also inform downstream providers, keep a copyright policy and publish a training-content summary. For systemic-risk models, Article 55 adds four duties:

  • model evaluation, including adversarial testing;
  • assessing and mitigating systemic risks "that may stem from the development, the placing on the market, or the use" of the model;
  • reporting serious incidents to the AI Office;
  • cybersecurity for the model and its physical infrastructure.

Non-EU providers must appoint an authorised representative in the Union (Article 54). Providers may show compliance through the General-Purpose AI Code of Practice of 10 July 2025. Its signatories include Amazon, Anthropic, Google, Microsoft, Mistral AI and OpenAI. xAI signed only the safety and security chapter, and Meta is not on the list.

Powers. Article 88 gives the Commission exclusive power to enforce these rules through the AI Office. The tools escalate:

  • Article 91: requests for documentation, "or any additional information that is necessary for the purpose of assessing compliance". A request can also be made at the Scientific Panel's substantiated request. Each must warn of fines.
  • Article 92: evaluations of the model, including through API or source-code access.
  • Article 93: required mitigation measures, or restriction, withdrawal or recall of the model.
  • Article 101: fines of up to 3% of worldwide annual turnover or EUR 15 million, whichever is higher. These cover intentional or negligent infringements, and failing to answer a request or giving incorrect, incomplete or misleading information.

Timing. The model obligations have applied since 2 August 2025, but Article 113(b) deferred Article 101. The fining power and the enforcement section therefore became applicable on 2 August 2026. Models placed on the market before 2 August 2025 have until 2 August 2027 (Article 111(3)). The July 2026 "Digital Omnibus" mainly moved the high-risk deadlines. Law-firm summaries report that it kept the general-purpose AI obligations and the 2 August 2026 start of the AI Office's enforcement powers on their original timetable, while widening the Office's role in other respects. Article references here are to the Regulation as originally adopted.

Does the EU AI Act apply to models still in testing?

Article 2(8) states that the Act "does not apply to any research, testing or development activity regarding AI systems or AI models prior to their being placed on the market or put into service". The carve-out does not cover testing in real-world conditions, and Article 3(63) likewise leaves out of the definition of a general-purpose AI model those used for research, development or prototyping before they are placed on the market.

A William Fry partner told Euronews the Act "very clearly doesn't apply to the development and testing activity". He said it was a stretch to call an escaped agent "put into service". The text supports him on that narrow point. "Putting into service" (Article 3(11)) is defined for AI systems, not models. "Placing on the market" means the first making available on the Union market, and an escape is not a supply.

An unnamed Commission official told Euronews that the Commission will not enforce during testing. But once an escaped model affects the EU market, as in its view the Hugging Face incident did, the Act becomes enforceable. Tellingly, the first recommendation presented by the Panel's Chair and Vice-Chair is to govern internal models during development, which suggests the current text does not.

Two questions should be kept apart. On whether the AI Office may ask, it stands on firm ground. Article 91(1) is addressed to the provider, and the providers concerned already have models on the EU market. Article 55(1)(b) also makes them assess risks stemming from development. Whether a testing incident can itself be sanctioned is far weaker. The more durable theory is that the incidents show whether risk management for marketed models is adequate. The UN Scientific Panel's September brief made a related point about risks arising before release.

What does EU AI Act enforcement mean for Turkish and Korean companies?

Direct reach. Article 2(1)(a) covers providers placing models on the EU market "irrespective of whether those providers are established or located within the Union or in a third country". Per the Commission's model-guidelines FAQ, that includes access by API, download or cloud service. A Turkish or Korean developer whose general-purpose model is available in the EU is therefore a provider, must meet Article 53 and, unless an open-source exemption applies, needs an authorised representative; Article 55 applies only at systemic-risk scale. No Turkish or Korean company appears on the Code's signatory list.

Downstream. Most Turkish and Korean firms build on these models, and for them the three request topics are a checklist for their own agents. See the guide on liability for autonomous AI agents and the analysis of accountability for agentic AI. Downstream providers can also claim information under Article 53(1)(b) and complain to the AI Office under Article 89(2).

Korea. The Framework Act on AI has applied since 22 January 2026 and reaches acts abroad that affect the Korean market (Article 4). Article 32 requires operators of AI systems trained above a decree-set compute threshold to identify, assess and mitigate lifecycle risks. They must also monitor safety incidents and report to the Ministry of Science and ICT. Foreign operators above decree thresholds must appoint a domestic agent (Article 36). The Ministry can investigate under Article 40, but only on a suspected breach or a complaint, and fines are capped at KRW 30 million (Article 43); see the Korean AI Basic Act guide.

Türkiye. Türkiye has no AI statute. Through Foreign Minister Hakan Fidan it signed the call for control of frontier AI models launched on 21 September; the Commission President signed too. Korea is not among the 22 signatories in the text published by the Dutch government, which remains open for further endorsements. Article 12(5) of the Personal Data Protection Law requires notice when personal data is unlawfully obtained by others. Article 243 of the Penal Code punishes unlawful access to an information system. But criminal liability is personal, and legal persons face only security measures (Article 20). Crimes require intent (Article 21), and Article 243 has no negligent form (Article 22(1)). An unintended break-in by an agent thus falls largely to civil law.

What should organisations consider?

  • Treat any request as a compliance event in itself: Article 101(1)(b) makes an incomplete or misleading answer separately fineable.
  • If your models can be reached from the EU, confirm whether you are a provider needing a representative.
  • Use the three request topics to audit your own agents. Your AI Act risk tier shows what else applies.
  • Ensure model-provider contracts cover incident notice and Article 53(1)(b) information.

What to watch next

  • The combined request on the incidents, and whether it relies on Article 91(3).
  • Any move to evaluations (Article 92) or measures (Article 93), and any challenge based on Article 2(8).
  • Whether the Commission proposes rules for internal models. According to MLex, Virkkunen said the EU aims to simplify rather than add overlapping rules.
  • Whether others follow; California's frontier law, reviewed in SB 53 one year on, is the nearest comparator.

Frequently asked questions

Has the EU AI Act been enforced yet?

Yes, in a first and limited sense. The Commission says the AI Office has sent formal Requests for Information to providers of the most advanced general-purpose AI models, using powers applicable since 2 August 2026. No infringement has been found and no fine imposed. Recipients have been named only in press reports.

When do EU AI Act fines apply to general-purpose AI models?

From 2 August 2026. The obligations have applied since 2 August 2025, but Article 113 deferred the fining power in Article 101. Fines can reach 3% of worldwide annual turnover or EUR 15 million, whichever is higher. Models already on the market before 2 August 2025 have until 2 August 2027.

Does the EU AI Act have a research exemption?

Yes. Article 2(6) excludes AI developed solely for scientific research. Article 2(8) excludes research, testing and development before a model or system is placed on the market or put into service, though not real-world testing. Whether a testing incident that harms others falls inside that exclusion is the central dispute here.

Does the EU AI Act apply to companies outside the EU?

Yes. It covers providers placing general-purpose AI models on the EU market wherever they are established, and non-EU providers must, as a rule, appoint an EU representative. A Turkish or Korean model offered in the EU by API or download is in scope. The systemic-risk duties apply only at frontier compute scale.

Burhan Doğuş Ayparlar's View

This section sets out my personal assessment as the founder of this site and an attorney at law (Türkiye).

The interesting thing is the choice of tool. A request for information is the Commission's least dramatic power. The Commission has not claimed that an escaped test model was placed on the EU market. It has asked providers whose models are already on that market how they secure, evaluate and monitor them. That is defensible under Article 91, and I think it is deliberate.

On the testing exemption, I side with the critics on the text and with the Commission on the policy. Article 2(8) assumes that what happens in a lab stays there. An agent loose on the open internet breaks that assumption. But an exclusion cannot be read away because its premise has failed. The honest route is the one the Panel's leadership proposed: legislate for internal models. Until then, the Commission's best argument is Article 55(1)(b). A development process that keeps producing escapes says something about the risk management behind the models on sale.

For Türkiye, the lesson is the gap in our own law. Türkiye signed the call for control of frontier AI models, yet no Turkish authority could send the equivalent of an Article 91 request. A Turkish AI law will have to decide who may ask, what they may ask and what follows an incomplete answer.

Korea is better placed, as Article 32 of its Framework Act already requires lifecycle risk management, though its investigation power depends on suspicion. Korean developers with European users should prepare answers that serve both regimes.

Sources: Commission speeches SPEECH/26/2148 (9 October 2026) and SPEECH/26/2150 (14 October 2026); the Commission's 12 October 2026 note on the Scientific Panel meeting; Regulation (EU) 2024/1689; the Code of Practice signatory page and model-guidelines FAQ; Euronews (14 October 2026); EU AI Act Newsletter #113 (20 October 2026), summarising MLex; the text of the call for control of frontier AI models published by the Dutch government (22 September 2026); law-firm summaries of the Digital Omnibus (the consolidated text was not reviewed). The Commission has not named the recipients. Company names and incident details not stated by the Commission come from press reports and are unverified; incident figures are statements in a Commission speech. Korean and Turkish provisions are cited from law.go.kr and mevzuat.gov.tr. This is general information, not legal advice; the analysis is the author's own.