U.S. export ban on Anthropic models jolts allies

U.S. officials cut foreign access to Anthropic’s Mythos 5 and Fable 5 models in June, forcing governments and companies in Europe, Canada and the U.K. to reassess their reliance on American frontier AI systems.

The order placed export controls on two cybersecurity-focused models from Anthropic. Anthropic had expanded Mythos access to 150 global organizations before releasing Fable 5, which added safeguards for cyber use. U.S. officials said researchers had jailbroken the model and argued that foreign access posed a national security risk.

The ban hit agencies, hospitals, research labs and private companies that had built projects around Anthropic access. Leaders in France, the Netherlands, Canada and the U.K. said the cutoff showed how much control Washington holds over AI infrastructure.

French Member of European Parliament Christophe Grudler said the U.S. held a “kill-switch” over key technology. Dutch Member of European Parliament Bart Groothuis urged Europe to build its own large language models and open-weight systems. Canadian Prime Minister Mark Carney said Canada should learn from the cutoff and diversify its AI supply.

The technical concern centers on cybersecurity automation. Anthropic marketed Mythos as a model that could find flaws in old codebases and help defenders harden software. Fable 5 offered comparable cyber capability with extra controls meant to limit abuse.

Those controls did not satisfy U.S. officials. They blocked export access by June 12, leaving foreign users without the model and forcing companies to seek alternatives. Nvidia and Adobe leaders have urged the Trump administration to restore access, arguing that defenders need advanced models to find vulnerabilities before attackers do.

The supply chain lesson looks familiar to chip buyers. Companies treat advanced AI models like strategic inputs, much as cloud providers and device makers treat GPUs, high-bandwidth memory and leading-edge wafers. A customer can pay for access, integrate the tool and train staff around it, yet a government decision can still sever access.

That risk changes procurement. Chief information security officers now have reason to sign backup contracts with non-U.S. model vendors, especially providers that offer open weights or local deployment. Open-weight models give customers more control because teams can run them on their own infrastructure, although they still need GPUs, power, talent and security staff.

Europe has pushed sovereign AI for years, but the Anthropic cutoff gives officials a concrete case. Governments can point to interrupted pilots in research, defense and health care. They can also justify spending on domestic compute clusters, data centers and model developers.

France has linked the Anthropic dispute to a wider shift away from U.S. technology vendors. French agencies have explored domestic alternatives to Palantir, messaging apps and search tools such as Qwant. European institutions have also studied open-source office software and cloud providers outside Amazon Web Services, Google Cloud and Microsoft Azure.

India could gain from that shift. French President Emmanuel Macron has promoted a joint French-Indian AI effort built around trusted and open systems. Such partnerships would let European governments reduce dependence on U.S. suppliers without trying to fund the entire AI stack alone.

The U.S. still holds a large lead. American companies control many of the strongest closed models, and Nvidia controls the accelerator market that trains and serves them. Chinese companies also have deep talent pools, national backing and large domestic markets. Europe and Canada can reduce risk, but they need years of investment to match frontier model labs.

The Fable 5 ban may also split the AI market by jurisdiction. U.S. vendors could keep their best models inside approved regions, while allies seek open systems they can audit and host. That would give customers more control but less shared scale. It would also force model makers to design products for export rules as much as benchmarks.

For chipmakers and cloud providers, the policy signal matters as much as the model cutoff. Buyers now have to price sovereign access into AI contracts. A model’s benchmark score, context window and tool use no longer settle the purchase decision. Customers will ask who can turn access off, which laws apply and which alternative can keep a project running.

The immediate fight concerns Anthropic’s Mythos 5 and Fable 5. The larger market shift concerns trust. Governments that depend on foreign frontier models now know they need compute, models and contracts they can control before the next export order arrives.