AI sovereignty risk is now real

A US government order shut down Anthropic’s Claude Fable 5 globally within hours — and changed the AI vendor risk conversation for every business outside the United States.

Top stories for week 25: 

  • The US government issued an export control directive ordering Anthropic to stop serving Claude Fable 5 and Mythos-5 to any foreign national, taking both models offline worldwide within hours.
  • Microsoft had already told its own staff not to use Claude Fable 5 forty-eight hours before the government acted
  • Both Anthropic and OpenAI are now formally on the path to public markets, with confidential S-1 filings submitted within a week of each other — meaning AI vendor risk disclosures are about to improve significantly.
  • Mistral rebranded its Le Chat assistant as Vibe and launched new pricing tiers, positioning itself as a European-headquartered alternative to US frontier models. 

For the first time in the history of commercial AI, a frontier model was pulled offline by government order. Anthropic’s Claude Fable 5 — available worldwide on Friday afternoon — was gone for every foreign national by Friday evening. 

The trigger was an export control directive citing national security. The scope is broad: it covers anyone without US citizenship, regardless of where they are sitting or who employs them. A foreign national working in San Francisco on a US visa is covered. So is every customer in London, Berlin, Stockholm, or Oslo. Anthropic had no mechanism to verify citizenship from an API call, so they shut Claude Fable 5 and Mythos-5 down for everyone, globally, within hours of receiving the letter. 

By Saturday, the European Commission was using the word “sovereignty”. 

For your business, this creates a third category of AI vendor risk. Alongside technical risk and commercial risk, there is now geopolitical risk — the risk that a model you depend on can be switched off not by the vendor, but by the government of the country the vendor is headquartered in. Switching cloud regions or API endpoints inside the same country does not hedge this risk. If the order applies to the lab, it applies to every cloud and every reseller the lab uses. 

The Microsoft footnote matters here. Microsoft told its own staff not to use Claude Fable 5 forty-eight hours before the US government did the same thing on a global scale. Even the cloud reseller hosting the model did not trust it for its own internal work. That is a meaningful early indicator for any governance team tracking what is coming. 

Your weekly briefing on AI in business