EU Unveils New AI Security Strategy as US Restricts Access to Frontier AI Models

The European Union risks losing access to the latest artificial intelligence models after the Donald Trump administration restricted their use to US companies and organisations designated as “trusted partners”. The new rules apply to leading AI systems, including Anthropic’s Mythos 5 and OpenAI’s ChatGPT 5.6, a move that could significantly undermine the competitiveness of European businesses in the AI sector.

The decision once again highlights Europe’s technological gap with the United States. Europe’s most advanced AI developer, the French company Mistral, still lags well behind its American competitors. In response, the European AI Forum, which brings together national AI organisations from across Europe, has called for public investment in artificial intelligence to be tripled and for public authorities to prioritise the procurement of genuinely European AI solutions.

Against the backdrop of restricted access to frontier AI models, the European Commission is preparing to unveil its first comprehensive strategy on cybersecurity and artificial intelligence. The initiative is expected to establish a framework for cooperation between Brussels and EU member states on assessing risks and ensuring the safe deployment of advanced AI systems. AI safety has become an increasingly pressing issue amid concerns that the most powerful models could facilitate large-scale cyberattacks or eventually operate beyond meaningful human control.

Particular attention has recently focused on Anthropic’s Mythos model, which the company initially released, then temporarily withdrew, before relaunching it after further development. Its unprecedented cyber capabilities have become one of the strongest arguments for strengthening international oversight of frontier AI models.

In response, European countries are rapidly expanding their AI safety infrastructure. After the United Kingdom became the first country in the world to establish an AI Security Institute in 2023, France launched its own institute, Germany announced plans to create a similar body, Estonia began developing an equivalent competence centre, and the Netherlands is considering establishing one as well. The UK institute has already secured agreements with leading US AI laboratories, allowing it to assess new models before their public release.

The European Union has also established its own supervisory body for artificial intelligence — the AI Office, created under the AI Act. Unlike the British institute, however, the AI Office combines safety oversight with regulatory enforcement and has the authority to impose penalties for non-compliance. According to experts, this dual role may discourage major AI companies from voluntarily providing early access to their models for pre-deployment testing. Analysts argue that US AI developers are more willing to cooperate with the UK institute precisely because it evaluates AI risks rather than policing compliance.

The European Union Agency for Cybersecurity (ENISA) has already begun testing OpenAI’s GPT-5.5 Cyber model but does not currently have access to Anthropic’s Mythos. At the end of June, the European Commission said ENISA, together with member states and other EU bodies, was developing a comprehensive framework for assessing the safety of frontier AI systems. According to Commission officials, the upcoming strategy will focus on coordinating national resources to ensure the safe deployment of Europe’s most advanced AI technologies.

Despite concerns about overlapping responsibilities, most experts support the creation of national AI Security Institutes. They argue that many AI-related threats fall within the domain of national security, which remains the responsibility of individual states, while national governments are also better positioned to compete for highly skilled AI experts through their own security budgets. Experts further suggest that countries could specialise in different areas of AI risk assessment — for example, Germany on cybersecurity threats, France on nuclear-related risks, and other member states on additional critical areas. Under such a model, the European Union’s primary role would be to coordinate national efforts rather than centralise all responsibilities.

Source: EURACTIV