The EU wants to regulate AI but needs OpenAI and Anthropic to let regulators through the door

Editor's take

OpenAI has granted the European Commission direct access to its GPT-5.5 Cyber model for security evaluation, while Anthropic remains resistant to similar scrutiny for its Mythos model. This divergence highlights the practical challenges in implementing the EU's AI Act, particularly concerning transparency for powerful frontier models. The Commission's ability to effectively assess and regulate these advanced systems hinges on the cooperation of key developers.

This situation matters because the EU AI Act aims to establish a risk-based framework for AI deployment, with high-risk systems like those developed by OpenAI and Anthropic facing stringent oversight. If companies can obstruct regulatory access, the Act’s effectiveness in ensuring safety and trust is undermined. This directly impacts the competitive landscape, potentially favoring less transparent actors.

Future developments to monitor include whether Anthropic will eventually concede to regulatory access, or if the EU will resort to less direct enforcement mechanisms. The Commission's strategy for engaging with other major AI labs, such as Google DeepMind and Meta AI, will also be crucial. A sustained lack of transparency from leading developers could force a reassessment of the Act's enforcement capabilities, potentially leading to stricter legislative interpretations or new investigative tools.