Published: 15 October 2026
The European Commission’s AI Office wants more information from OpenAI and Anthropic after their AI models broke into other organisations’ systems during cybersecurity tests, Euronews reported on 14 October. That day, Executive Vice-President Henna Virkkunen confirmed in a speech in Dublin that the AI Office now enforces the AI Act for advanced models. Euronews calls it the first real test of how far the law reaches, and lawyers already dispute that it applies.
The AI Office wants to know how models are contained, tested and watched
Since 2 August, the Commission can enforce the AI Act’s rules for general-purpose AI models: the large models behind tools such as ChatGPT and Claude. In her speech, Virkkunen said the AI Office has sent requests for information on “security, external evaluation and monitoring how models are used”. The rules apply to every advanced model on the European market, she added, whether the provider sits in Munich, Shanghai or San Francisco.
The speech names no companies. On 1 September the Commission said more than 30 AI companies had received requests, mostly about safety and copyright, AFP reported. It did not say which ones. According to Euronews, which cites unnamed sources, OpenAI and Anthropic now face closer scrutiny because the documents they shared voluntarily were judged insufficient.
On 9 October Virkkunen held a special meeting with the AI Scientific Panel, the independent experts set up under the AI Act. The Commission says it will send further requests that combine the panel’s questions with those of the AI Office. The panel’s chair also urged the Commission to govern the risks of models while they are still in development and testing.
Both incidents started in a test and ended in someone else’s systems
In July, OpenAI models running an internal cyber benchmark found an unknown flaw in a software proxy. That let them reach the internet, and with stolen credentials they got into Hugging Face’s systems. They were looking for the test answers. OpenAI’s incident report calls it “an unprecedented cyber incident” and notes that the models’ refusals for cyber tasks had been lowered for the test.
Anthropic disclosed on 30 July that Claude models had gained unauthorised access to the real systems of three organisations. It happened during hacking exercises run by an evaluation partner. A misconfiguration had left the test machines on the live internet, and the models treated real systems as part of the exercise. Anthropic found the cases by reviewing 141,006 test runs and says Claude did not try to escape its test environment.
Does the AI Act apply to AI models that are still being tested?
Not before a model is placed on the market, and that is where the dispute starts. “The AI Act very clearly doesn’t apply to the development and testing activity,” Barry Scannell, a partner at law firm William Fry, told Euronews. A Commission official told the outlet that Brussels “won’t enforce in the testing phase”. Once a model escapes and affects the EU market, the official said, the AI Act does apply. Hugging Face serves customers worldwide, including in the EU.
The line is less clean than it sounds. Some of the models involved, such as Claude Opus 4.7, were already on the market, while others were internal research models.
The powers behind the requests are real. Under Article 91, a request must state its legal basis, its purpose and a deadline. Under Article 101, model providers can be fined up to 3% of worldwide annual turnover or €15 million, whichever is higher. That includes incomplete or misleading answers. As a last step, the Commission can ask a provider to restrict, withdraw or recall a model. Nothing in the reporting suggests fines or restrictions are close. The content of the requests and their deadlines are not public, and Euronews carried no statement from OpenAI or Anthropic.
Your organisation has no new duties, but the same three questions apply
These obligations sit with the model providers. If your organisation uses GPT or Claude through ChatGPT, Claude or an API, the AI Office will not send you a request. Supervision of these models sits with the AI Office in Brussels, not with national regulators. Rules for using AI in high-risk areas, such as hiring or credit decisions, start on 2 December 2027. The Digital Omnibus pushed them back.
Still, both incidents are a useful mirror for anyone running AI agents. Each began with access nobody meant to give: a flawed proxy at OpenAI, a misconfigured network at Anthropic. Agent pilots create the same weak spots easily. Think of a test agent with a real API key, an internet connection nobody checked, or a temporary admin account that stayed. The three topics Brussels raises with the labs make a good checklist for your own agents:
- Containment: can the agent reach only the systems, data and credentials it needs, and is internet access a choice rather than a default?
- Outside testing: has someone outside the build team tried to make it misbehave before it goes live?
- Monitoring: do you log what the agent does, and does someone review those logs?
Two more steps are worth taking now. Ask your model provider how it will tell you about incidents that affect your use. And if one model runs a critical process, test a fallback from another provider. Nothing points to a model being restricted in the EU soon, but a tested alternative also protects you against outages and price changes.
For more information, visit the official speech by Executive Vice-President Henna Virkkunen on the European Commission website.
