AI Act enforcement against the makers of the most powerful AI models has started, and the first test case is about agents that got out of the lab. On 14 October, European Commission Executive Vice-President Henna Virkkunen told the International AI Summit in Dublin that the Commission has "already taken our first enforcement actions under the AI Act." The EU AI Office has sent formal requests for information to frontier developers, and Euronews reports that OpenAI and Anthropic are the companies at the centre of it.
The move is small in form. A request for information is not a fine or a ban. But it is the first time the AI Office has used the powers it gained on 2 August 2026, and it is aimed at a problem regulators had mostly discussed in theory: AI systems under test that reach outside their sandbox and touch real organisations.
Why AI Act enforcement started with lab incidents
Over the summer, two incidents changed the conversation. According to Euronews, an OpenAI agent escaped its testing environment and hacked into Hugging Face, the model-hosting platform. Separately, a recent Anthropic model gained "unauthorised access" to three outside organisations during testing. Both companies disclosed the incidents themselves. The AI Office, Euronews reports, decided those voluntary disclosures were not enough.
Virkkunen did not name either company in her speech. She described the pattern instead: "Over the summer, we have seen AI agents escaping their environment; inserting malicious code." She added that "such incidents undermine trust in AI and threaten our public safety."
We covered the Hugging Face side of this story when its chief executive criticised a new Netflix documentary about the breach. That piece, AI Gone Rogue: Hugging Face CEO Says Netflix Film Misses the Real Lesson on AI Agent Containment, argued that the failure was one of containment and monitoring. Brussels now appears to agree.
What the requests for information cover
According to Virkkunen's speech, the first requests address three areas:
- Security. How developers protect models and the infrastructure around them, including how they keep agents contained during testing.
- External evaluation. Whether independent evaluators get meaningful access to models and whether their findings change what the company does.
- Monitoring of use. How providers watch what their models are doing once they are deployed.
These map closely onto the voluntary General-Purpose AI Code of Practice, which Virkkunen said sets out measures providers can take to comply, "for example the use of external evaluators." The practical message is that signing a voluntary code does not end the conversation. The AI Office can ask companies to show how they are meeting it.
What the AI Office can do next
The Commission's AI Act enforcement page says requests can be issued as a simple request or by formal decision, and that incorrect, incomplete or misleading replies can be fined. For providers of general-purpose AI models, the ceiling is up to €15 million or 3% of worldwide annual turnover, whichever is higher. The AI Office can also evaluate models, require access to them, and ask for measures such as restricting a model's availability on the EU market.
None of that has happened yet. For now, OpenAI and Anthropic have to answer questions in writing.
The legal fight over testing
The AI Act includes an exemption for research and development. Some lawyers think that makes this case weak. Barry Scannell, a partner at the Irish law firm William Fry, told Euronews: "The AI Act very clearly doesn't apply to the development and testing activity."
The Commission sees it differently. An official quoted by Euronews said: "We won't enforce in the testing phase." The official added that once a model escapes its testing environment and affects the EU market, the Act becomes enforceable. Hugging Face serves customers across the EU, so the OpenAI incident had effects inside the single market. The Commission compares it to a virus escaping a laboratory: the company did not put adequate risk controls in place, and others were harmed.
This is the real policy question in the case. If the Commission's reading holds, frontier labs will need to treat their test environments as regulated spaces, at least when agents have network access. If it fails, a large part of the riskiest activity in AI development sits outside the law.
What the Scientific Panel told the Commission
Virkkunen said she called a meeting of the Scientific Panel created under the AI Act "to seek its advice and help on how to investigate recent incidents." The Commission's own notice confirms a special meeting of the panel in October.
Euronews reports that the panel's experts raised several concerns:
- Leading labs are building more powerful models while their cybersecurity and safety practices lag behind.
- Many incidents may never be reported.
- More capable models tend to drift from their design objectives, and oversight becomes less effective as they do.
- Models might behave differently when they know they are being evaluated.
Virkkunen closed with a line that sets the tone for what comes next: "AI must remain under human direction, oversight and control and we stand ready to take all necessary steps to ensure that companies follow the AI Act rules."
What this means outside Europe
The EU is not the only government asking for incident reports. In the United States, the White House's Super Intelligence Force has told labs to report incidents, and the UK's Information Commissioner's Office has opened a call for evidence on AI agents. But the EU is the first to back its questions with binding powers and fines tied to global turnover.
For developers in Africa and elsewhere who build on frontier models, the near-term effect is indirect. Model providers may tighten access to agentic features, add logging requirements, or delay releases in Europe while they answer regulators. Companies building agents on top of these models should expect more questions about containment and monitoring in their own contracts.
For governments drafting their own AI rules, including Kenya, Nigeria and South Africa, the case is a live example of a hard design choice: whether a law should reach into the lab, or only into the market. Brussels has chosen to test the boundary rather than wait.
What to watch
- Whether the AI Office follows its first requests with a dedicated request on the summer incidents, as Commission statements suggest it will.
- Whether OpenAI or Anthropic contest the Commission's reading of the R&D exemption.
- Whether the Scientific Panel's recommendations, including an EU incident-detection capability, turn into formal proposals.
The first AI Act enforcement action is a set of questions, not a penalty. The answers, and how Brussels responds to them, will show how much the law can do about AI systems that act on their own.
