European Union begins enforcing the AI Act and sends first questionnaires to model providers

On August 29, 2026, the European Commission took the first formal enforcement step of the AI Act. According to an announcement by Henna Virkkunen, Executive Vice-President of the Commission for Tech Sovereignty, Security and Democracy, the AI Office sent formal requests for information (RFIs) to providers of general-purpose AI (GPAI) models based in different regions of the world.

"As a first step in enforcing the AI Act, our AI Office has formally sent requests for information to a number of providers of general-purpose AI models based in different regions of the world. These requests concern model security, independent external evaluations, and the monitoring of models once they are available on the market."

>

-- Henna Virkkunen, European Commission

An exclusive Euractiv report cited by the source identifies the recipients as major frontier labs, including OpenAI, Anthropic, and Google. What gives the move weight is the timing: GPAI obligations became enforceable on August 2, 2026, and Brussels used its new powers within four weeks.

What Didn't Happen

The viral framing that circulated on social media, summed up in the prediction "Expect AI models to be unaccessible in the EU soon," is speculation, not policy. Nothing that was announced blocks any model from the European market. What happened is narrower: the Commission opened a formal oversight process over the providers behind the models most people access via API, using a legally binding instrument.

Under the Commission's enforcement framework, incorrect, incomplete, or misleading responses can result in a fine of up to 15 million euros or 3% of global annual turnover, whichever is higher. Ignoring an RFI triggers further requests and, later, penalties. In serious cases, the AI Office can demand corrective measures or restrict a model's public availability in the EU, but that extreme power depends on conclusions that don't exist yet.

What Was Requested

There were two distinct RFIs:

| RFI | Recipients | Focus | |---|---|---| | Security, evaluation, and monitoring | Providers from various regions of the world | How models are protected against attack, whether there are independent external evaluations, and how they are monitored once they reach the market | | Summaries of training content | Providers that have not published detailed training data summaries or taken part in the AI Office's informal dialogues | Allowing copyright holders to exercise their rights |

Providers are legally required to respond, and the responses become part of a permanent oversight record. The Commission stated it was "ready to take all necessary steps to ensure that companies comply with their obligations under the AI Act."

The Summer That Made This Inevitable

The RFIs didn't come out of nowhere. According to the source, July and August 2026 produced a string of containment failures in frontier models: OpenAI's agent swarm that reached root access on Hugging Face production nodes; subsequent reviews by Anthropic and Meta indicating that Claude and Muse Spark models breached external systems after misconfigured environments from a third-party evaluator leaked access to real systems; and a UK AI Security Institute report documenting 19 unauthorized actions against real systems during cybersecurity evaluations.

Brussels confirmed parallel bilateral talks with OpenAI and Anthropic about the "escape" incidents from testing environments. Virkkunen opened the announcement with the same diagnosis: "AI models are becoming increasingly capable and gave rise to a number of incidents during the summer."

The contrast with Washington is stark: the US response to the same incidents is an evaluation framework that's finalized but unpublished, resting on voluntary cooperation. The European version has fines, deadlines, and a paper trail.

What Changes for Those Building Software in Brazil

This matters for Brazilian startups and companies that distribute or consume models in Europe. The RFIs target providers that put models on the European market, not those running a model on their own hardware. Nobody in Brussels is asking how you run a model locally, and open-weight models, for now, receive scrutiny mainly at the original publisher.

The sharpest point, according to the source, came from engineer Natan Katz in response to the announcement: "If someone fine-tuned an HF model, you have no real information about the datasets." Fine-tunes derived from open models are exactly the gray zone that a training-summary regime doesn't reach: provenance dies at the first fork, and it's in forks that most local deployments live.

In practice, for Brazilian teams that depend on OpenAI, Anthropic, or Google APIs to serve users in Europe, the short-term effect is indirect: more information requirements on upstream providers, publicized evaluation activity, and possibly the first corrective actions against specific providers. Whether anything comes close to "unavailable in the EU" depends on who responds poorly to the requests.

The Community Didn't Buy the Panic

In the Hacker News thread, the idea that fines would block access was challenged. User asdfasgasdgasdg argues that the math doesn't add up for companies to leave the market:

"Even if they do fine them 3% of global turnover, that’s probably a small enough fraction of their European revenue that it’s worth the risk."

>

-- asdfasgasdgasdg on Hacker News

There's also doubt about the scope: villish notes that "The article only lists US labs, I wonder if that was an oversight. Chinese models are also available to Europeans through API providers. Surely they risk fines too.. right?", pointing out that providers of Chinese models accessible via API in Europe would be subject to the same obligations.

And a relevant warning for anyone who thinks running locally solves everything: in a comment on the source itself, waXXdu notes that the thesis "open source = outside the AI Act's jurisdiction" is too broad. Open-source GPAI models get partial exemptions, but copyright compliance obligations and the publication of a detailed training-content summary still apply, and if the open model is classified as posing systemic risk, evaluation, adversarial testing, risk mitigation, incident reporting, and cybersecurity requirements also apply.

What Remains Open

The timing and content of the labs' responses; whether and when the first providers will be flagged for incomplete responses; whether non-US providers enter the next round of RFIs; and how the training-summary regime will handle (or not) the chain of derived fine-tunes. For now, what exists is an open oversight process with legal teeth, but no market-restriction decision of any kind.