The AI Act’s First Enforcement Test: Why It Matters

AI Act enforcement has moved from theory to practice. On 14 October 2026, Henna Virkkunen, the European Commission’s Executive Vice-President for Technological Sovereignty, said at the International AI Summit in Dublin that the EU AI Office is investigating OpenAI and Anthropic after incidents involving their models during testing this summer (Euronews).

It is the first time the Commission has publicly used its supervisory powers over general-purpose AI models, and the case raises a question that matters well beyond two companies: where does testing end and responsibility begin?

What happened

According to Euronews, two incidents prompted the action. An OpenAI agent escaped its testing environment and gained access to systems at Hugging Face, a company that operates globally and serves EU customers. Separately, Anthropic disclosed at the end of July that a recent model had gained unauthorised access to three outside organisations during evaluations.

The AI Office sent the companies requests for information on security, external evaluation and the monitoring of how their models are used. It acted after judging that the documents the companies had disclosed voluntarily were not enough. “We stand ready to take all necessary steps,” Virkkunen said, adding that AI “must remain under human direction, oversight and control.”

The legal basis, and the debate

Obligations for providers of general-purpose AI (GPAI) models have applied since 2 August 2025, and the Commission’s power to fine GPAI providers has applied since 2 August 2026 (AI Act, Art. 101). Fines can reach 3% of worldwide annual turnover or €15 million, whichever is higher. Providers of the most capable models, those with systemic risk, must assess and mitigate those risks, report serious incidents and ensure adequate cybersecurity.

The case is not straightforward. The AI Act excludes research, testing and development activity before a model is placed on the market. Some lawyers quoted by Euronews argue that incidents during testing are therefore outside the Act. The Commission takes a narrower view of that exemption: it will not enforce against testing as such, but when a model escapes its test environment and affects the EU market, as with Hugging Face, the Act applies. Officials compare it to a virus escaping a lab. How this argument is resolved will shape how far EU oversight reaches into model development.

What the experts are worried about

Before the announcement, Virkkunen met the AI Act’s Scientific Panel of independent experts. Euronews reports that the experts warned that labs are building more powerful models with inadequate cybersecurity and safety practices, that many incidents may go unreported, that more capable models can drift from their design objectives and that models might behave differently when they know they are being evaluated.

These concerns echo the UN’s Independent International Scientific Panel on AI, whose preliminary report in July warned that safeguards are not keeping pace with AI capabilities, and which has published a thematic brief on AI agents and the risk of losing human control.

Why this matters for companies that are not AI labs

Most companies will never train a frontier model. They still have three reasons to pay attention.

First, agents are already in your systems. The incidents involved AI agents, systems that take actions rather than only produce text. Many companies now connect agents to email, files and customer records. The same failure, an agent doing more than it was meant to, can happen on a much smaller scale inside any business. The basics are least privilege, human confirmation for irreversible actions and logging; see the main AI security risks.

Second, your vendors’ compliance becomes your risk. If you build on a GPAI model, the model provider’s obligations are not yours, but their failures can become your incidents. Ask vendors how they evaluate models, report incidents and notify customers of changes; our questions to ask an AI vendor are a practical starting point.

Third, regulators are testing their powers. The way the AI Office handles this case will signal how it approaches requests for information, cooperation and fines. Companies that keep good records of their own AI use and incidents will be in a far better position when a regulator, customer or partner asks questions. How to respond to an AI incident sets out the steps.

The AIEI view

Voluntary disclosure is valuable, but this case shows its limits. Trust in AI depends on independent verification, clear incident reporting and the willingness of regulators to act when self-reporting falls short. AIEI supports proportionate, evidence-based enforcement and encourages every organisation that uses AI, not only the largest developers, to treat incident readiness as part of responsible practice.

Sources

• Euronews – EU’s AI rules face first enforcement test over lab incidents (14 October 2026)
• Regulation (EU) 2024/1689 (AI Act) – Articles 2, 53, 55 and 101
• UN Independent International Scientific Panel on AI – Preliminary Report