OpenAI and Anthropic: Brussels Faces the Challenge of Access to AI Models

The EU and Access to Advanced AI Models

The European Union, equipped with the AI Act and the AI Office, faces a major obstacle: access to advanced artificial intelligence models. OpenAI recently reached out to the European Commission by offering access to its model, GPT-5.5 Cyber. This offer, confirmed by Thomas Regnier, spokesperson for the Commission, is seen as a welcome gesture of transparency. It would allow regulators to monitor the deployment of the model and address security concerns.

However, this proposal highlights an uncomfortable truth for Brussels: European regulators depend on AI companies to voluntarily grant access to these technologies. Discussions between OpenAI and the EU are already underway and are expected to continue this week. However, it remains to be determined which entities within the EU will benefit from this access, with names like ENISA, the AI Office, and DG Connect among the potential candidates.

Regnier took the opportunity at the conference to emphasize the importance of existing and forthcoming EU legislation. He mentioned that the current situation demonstrates the relevance of existing laws, such as the AI Act and the Cyber Resilience Act. He also advocated for a swift advancement of the Cybersecurity Act in the legislative process.

More Complex Negotiations with Anthropic

In contrast to OpenAI, talks with Anthropic are proving to be more complicated. Although the Commission has held "four to five meetings" regarding Anthropic's Mythos model, progress is slow. This model has sparked the current debate on cybersecurity. Regnier noted that, unlike OpenAI, Anthropic has not yet offered direct access to its model.

When asked about a potential request for access to Mythos, Regnier avoided answering directly, indicating that the access offered by OpenAI would be an ideal model for discussions with Anthropic.

Security and Regulatory Challenges

The issue of security remains crucial. Politico raised concerns about the security measures the Commission could implement if access to the models were granted. Regnier assured that precautions would be taken but did not provide specific details.

Currently, the EU has little information on Anthropic's Claude Mythos Preview, which remains reserved for select technology partners under the "Project Glasswing." Only the British Institute for AI Security has had the opportunity to test Mythos directly.

This situation highlights Europe's dependence on American AI companies, underscoring a power imbalance that complicates regulatory efforts.