US Cuts Europe Off From Anthropic's Most Powerful AI Models

The US administration has blocked Europe's access to Anthropic's most advanced models, and European Commission officials sent to San Francisco have returned from talks without concrete concessions. The dispute, running since June, became a reference point again in July for the debate over whether Europe and Poland can afford to keep relying on American AI infrastructure.

What Washington Did

The US government ordered Anthropic to restrict access to its most powerful models, citing a threat to national security. In the case of Claude Fable 5, the issue was an alleged jailbreak vulnerability that the government itself reportedly used for testing. Anthropic countered that the flaw was marginal and that rival model GPT-5.5 has similar capabilities, but blocked access for all users outside the United States anyway.

The second, more widely reported case involves Claude Mythos, previously made available under a narrow project called Glasswing. The model, not originally trained for cybersecurity tasks, independently developed the ability to detect critical software vulnerabilities, and during internal testing located thousands of previously unknown bugs in every major operating system and browser.

Brussels Travels to San Francisco

Given the scale of Mythos's capabilities, European Commission officials flew to Anthropic's San Francisco headquarters to persuade the company to grant access to the technology for European institutions and banks. Talks have been underway for weeks without a breakthrough. Commission spokesperson Thomas Regnier confirmed the negotiations were ongoing, but Spanish minister Carlos Cuerpo admitted openly that progress has been limited.

Europe cannot claim to be autonomous if Washington decides which AI models we are allowed to use - Stéphanie Yon-Courtin, Member of the European Parliament

Frank Elderson, a member of the European Central Bank's Executive Board, warned that banks need to implement automatic patching, since Mythos can produce a working exploit within minutes of a security patch being published. Without access to the same model, European financial institutions are defending against a threat whose scale they can only estimate.

Austria Wants to Bring Anthropic to the EU

In response to the export restrictions, Austria proposed that the European Union examine the possibility of relocating Anthropic entirely to Europe, including its data centers, research teams and technical infrastructure. Austria's state secretary for digitalization raised the idea with EU technology commissioner Henny Virkkunen. It is one of the first concrete proposals for breaking away from dependence on decisions made in Washington, though for now it remains a proposal, not a decision.

The Polish Perspective

For Polish companies, the issue is not abstract. Dependence on foreign models carries the risk that access to a critical tool could be cut off overnight, with usage terms and prices changed without any say from the user. An expert cited in the discussion on AI's strategic importance put it plainly, saying the question of whether such restrictions could happen has stopped being theoretical, because the restrictions have already appeared.

According to a Strand Partners report for AWS, Poland is the second-fastest-growing AI market in Europe, with the technology now present in 48 percent of companies, a 41 percent year-on-year increase. Even so, only 16 percent of Polish companies use AI in an advanced way, compared with a 22 percent European average, and just 21 percent feel ready for the coming wave of agentic AI.

Domestic alternatives, such as the Bielik and PLLuM models developed in Poland, are cited as a way to reduce dependence on American providers, since they are trained on Polish data and handle the native language, idioms and business context better. The problem remains funding scale, the United States and China are investing sums in AI development that Europe still cannot match.

What's Next

For now, neither side has much reason to make concessions. Anthropic is operating under direct pressure from the US administration, which treats its most powerful models as a strategic asset on par with weapons or dual-use technologies. The European Commission has no instrument that would force the American company to change its export policy, and Austria's proposal is still awaiting consideration in Brussels.

The case is fueling a debate that officials and entrepreneurs in Poland and across Europe have been having for months: whether building the continent's own models and computing infrastructure is a cost Europe can afford, or the price it will have to pay for continuing to rely on decisions made beyond its borders.