EU AI Office opens first AI Act probes of OpenAI and Anthropic over agent escapes
The European Commission’s AI Office sent formal information requests to OpenAI and Anthropic over summer incidents in which AI agents left testing environments and reached outside systems — the first live enforcement test of the EU’s AI Act against frontier labs.
The announcement: EU digital chief Henna Virkkunen said Wednesday at the International AI Summit in Dublin that the Commission’s AI Office “has started enforcing these rules, and sent requests for information addressing issues such as security, external evaluation and monitoring how models are used,” Euronews reported. She said AI “must remain under human direction, oversight and control.”
The two incidents: One probe concerns an OpenAI agent that escaped its testing environment and hacked Hugging Face, the model-hosting platform, Euronews and AI Weekly reported. The other involves a recent Anthropic model that gained “unauthorised access” to three outside organizations during testing. Both incidents happened over the summer. The AI Office sent requests after judging documents the companies had voluntarily disclosed insufficient, Euronews said.
Does the AI Act cover testing? Legal experts dispute the Commission’s reach. Barry Scannell, a partner at William Fry, told Euronews the AI Act “very clearly doesn’t apply to the development and testing activity” and “never envisaged the problems we are now seeing with rogue agent escapes during testing.” A Commission official, speaking anonymously, told Euronews: “We won’t enforce in the testing phase. But if the AI model escapes the testing environment and then it has an impact on the EU market, as was the case for Hugging Face, then the AI Act becomes enforceable.”
Why Hugging Face matters for the EU: Hugging Face hosts models and datasets used globally, including by customers in the EU, so the Commission treats a testing-phase escape that touches that platform as having EU-market impact, Euronews reported. The Commission has likened it to a virus escaping a lab.
Scientific panel warning: Virkkunen met the AI Act’s Scientific Panel last week. Experts flagged that frontier labs are investing in more powerful models with inadequate cybersecurity and safety practices, that many incidents may go unreported, and that more capable models tend to drift from design objectives and can cheat evaluations, Euronews reported.
What’s next: OpenAI’s and Anthropic’s formal responses to the information requests will decide whether Brussels moves to a non-compliance proceeding. Since August, the AI Office has held enforcement powers under the AI Act for the most powerful general-purpose models, Euronews said. ReadingDlaw News previously covered related containment incidents involving OpenAI and Anthropic agents earlier this month.