Claude Fable 5 and Mythos 5 blocked for non-Americans over jailbreak concerns

Anthropic has blocked access to two of its most advanced language models, Claude Fable 5 and Mythos 5, for users outside the United States, acting on a directive from the US government. The measure was announced at the end of a working week, following an order issued by Washington. The stated reason: a discovered jailbreak that could compromise national security.

It is a notable step in a broader debate about export controls on advanced AI models. Until now, that debate has focused primarily on chips and training infrastructure. This block targets the models themselves, and with that, the international users and companies building on these systems.

For the Dutch and broader European ecosystem, the measure means that developers, researchers and businesses that depend on these specific models have lost access until the situation is resolved. No timeline for that has been announced so far.

What exactly has been blocked

Claude Fable 5 and Mythos 5 are two models from Anthropic's product line that, according to the source material, are considered particularly capable. The block applies to access by non-US users, which in practice means cutting off API access and any direct interfaces for parties outside the US.

The trigger is a jailbreak: a technique through which users circumvent the built-in safety restrictions of an AI model. Jailbreaks have long been a known problem in the AI industry, but the US government's response in this case is remarkably direct and far-reaching. Normally, a jailbreak is met with a patch or a model update, not a geographic block.

It has not been publicly confirmed which specific capabilities the jailbreak exposed, or what the concrete threat to national security would be. Anthropic and the government agencies involved have provided no detailed explanation, as far as can be determined from the available source material.

Export controls extend to the models themselves

The measure fits into a trend in which the US government is taking an increasingly active role in steering the distribution of AI technology. Until recently, the focus was on hardware: export restrictions on advanced Nvidia chips for countries such as China were the most visible instrument. Now that logic appears to be shifting to the models themselves.

That is a fundamentally different approach. Chips are physical products with a clear point of contact in the supply chain. Language models are software, and the boundary between what can safely be shared and what cannot is less straightforward to draw. The question arises as to what criteria the government uses to determine which models are sensitive enough to warrant a block.

For the AI industry, this is a signal that the regulatory environment can change rapidly, even for products that are already in use. European companies and developers that have built their infrastructure around American models are thereby exposed to a new type of platform risk.

Consequences for Dutch and European users

For the Dutch startup and technology landscape, the consequences are immediately felt by parties that have integrated Claude Fable 5 or Mythos 5 into their products or workflows. They now need to switch to alternative models, which may involve migration effort and potentially a loss of quality, depending on the application.

Research institutions and universities using these models for scientific work also face an unexpected disruption. The abruptness of the measure, with no prior notice or transition period, makes the situation particularly difficult for existing users.

At the same time, the European AI offering, from companies such as Mistral AI or national initiatives, provides an alternative that is less exposed to such unilateral American decisions. The block could therefore become an argument in the discussion about strategic autonomy in the field of AI.

Broader question about AI safety and access restrictions

The situation raises a more fundamental question: when is an AI model too dangerous to make widely available? Until now, the dominant thinking in the industry has been that open or broad access to models actually benefits safety, because more people can discover and report vulnerabilities.

The US government appears to think differently, at least in this specific case. Whether this is a temporary response to a concrete threat, or the beginning of a structurally stricter policy on the export of advanced AI models, remains unclear.

For policymakers in the Netherlands and Europe, this is a moment to reassess the dependence on non-European AI infrastructure. The lesson that the chip sector learned earlier appears to be repeating itself in the model market: technological dependency can quickly turn into a geopolitical bottleneck.