Anthropic Cut Off by Washington: AI Dependent on the United States

A Sudden Cut Orchestrated by Washington

On June 12, a directive issued by the U.S. government forced Anthropic to suspend access to two of its artificial intelligence models, Fable 5 and Mythos 5, for all its clients worldwide. This decision, motivated by export control regulations, aimed to prohibit access to these models for any foreign national, whether inside or outside the United States. This measure had immediate repercussions for companies that rely on these models for their daily operations.

Highlighting Dependence on American Giants

This event sheds light on a risk often overlooked by businesses: their dependence on a limited number of AI suppliers, primarily American, such as OpenAI, Anthropic, Google, and xAI. These giants control a significant portion of the cutting-edge model market. A regulatory decision made in the United States can thus have global repercussions, often without warning for clients.

Anthropic: A Revealed Vulnerability

Anthropic emphasized that the U.S. directive was based on a minor vulnerability, already known and present in other public models, such as OpenAI's GPT-5.5. Despite rigorous security measures, the company had to comply with the directive, revealing the fragility of businesses in the face of foreign administrative decisions. If this standard were applied across the industry, it could paralyze many AI model deployments.

AI: An Essential Infrastructure

In just two years, generative AI has transitioned from being a gadget to an essential production tool. Companies use it for various tasks such as internal assistance, large-scale content generation, and code review. It has become an infrastructure as crucial as electricity or the cloud. However, unlike these infrastructures, AI often lacks redundancy. Most companies depend on a single supplier, and these suppliers are predominantly American. A sudden cut, due to a regulatory decision made halfway around the world, can therefore have disastrous consequences.

What the Anthropic Case Reveals

Anthropic is not a negligent player. The company recalls having subjected its safeguards to thousands of hours of red-teaming with the U.S. government, the British AISI, and third parties, and having adopted a defense-in-depth strategy, including imposing a 30-day data retention policy to detect and neutralize attacks. It claims that no "universal jailbreak" has been found. The issue is not the negligence of a supplier. The issue is that an administrative decision—based, according to Anthropic, on a narrow vulnerability widely available elsewhere—was sufficient to remove a tool used by hundreds of millions of people from the market within hours. If this standard were applied to the entire industry, it would freeze, by Anthropic's own admission, the majority of cutting-edge model deployments.

The Risk of AI Supply

The debate on AI risks has focused on internal issues: hallucinations, data leaks, GDPR compliance, and the AI Act. These are real concerns. However, the Anthropic case highlights an external danger that has rarely been anticipated: the supply risk. Export controls, sanctions, litigation, data localization requirements, regulatory reversals—these are all levers that can cut access to a model without any line of your code being at fault. For a European company, the leading supplier is almost always subject to foreign law. Your service continuity thus depends on a chain you do not control and have often not even mapped.

The Importance of Sovereign AI in Europe

In light of these challenges, the idea of sovereign AI makes perfect sense. In Europe, players like Mistral AI and infrastructures such as OVHcloud and Scaleway offer alternatives to American suppliers. Although the European ecosystem is more limited, it represents a viable option to ensure service continuity in the event of a cut from American models.

Measures to Mitigate Risk

To protect against such interruptions, companies should:

1. Abstract the Supplier: Design their infrastructure to be able to switch models without rewriting all their applications.
2. Keep a Ready Alternative: Integrate a second supplier, ideally European, to ensure a quick transition if needed.
3. Map Exposure: Identify critical processes dependent on each model to better anticipate the impacts of a cut.

In 2026, the crucial question for companies will no longer be just which model is the best, but who can control access to it and how to ensure resilience against unforeseen regulatory decisions.