On June 12, a US directive forced Anthropic to cut two models for all its customers. For businesses, the real question: who can turn off your AI, and do you have a plan B?
I do not run an AI research laboratory, but the marketing of a B2B SME. On a daily basis, our technical stack depends on models that we do not host. This June 12, 2026, the news highlighted a major flaw that most companies do not include in any risk matrix.
That day, Anthropic — one of three or four leading model providers in the world — announced that it had to suspend access to two of its models, Fable 5 and Mythos 5. Not because of an outage. Because of an export control directive from the American government, which prohibited access to these models to any foreign national, inside or outside the United States. To comply, the company had to cut them for all of its customers around the world. Its other models remain in service.
According to Anthropic, the directive would target a circumvention technique – a “jailbreak” – which the company considers minor, already known and reproducible on other public models, including OpenAI’s GPT-5.5. Anthropic disputes the decision, complies with it and says it is working to restore access. But for a steering committee, this technical debate is not the subject. The subject is what the episode reveals.
AI has become infrastructure — plugged into a single tap
In two years, generative AI has gone from gadget to production tool. Internal assistants, commercial co-pilots, large-scale content generation, code review: in many companies, real processes now run on these models. It is an infrastructure, just like electricity or the cloud.
However, a robust infrastructure is designed with redundancies. No one runs a factory on a single power line without backup. However, on the AI side, the majority of companies are connected to a single supplier — and these cutting-edge suppliers are concentrated, and overwhelmingly American: OpenAI, Anthropic, Google, xAI. The day one of them is cut, by a regulatory decision taken on the other side of the world, the customer learns it at the same time as everyone else. Without notice, without change.
What the Anthropic affair really says
Anthropic is not a careless actor. The company recalls having subjected its safeguards to thousands of hours of red-teaming with the American government, the British AISI and third parties, and having adopted a defense in depth strategy, even imposing a 30-day data retention to detect and neutralize attacks. She claims that no “universal jailbreak” has been found.
The point is therefore not the negligence of a supplier. The point is that an administrative decision — based, according to Anthropic, on a narrow vulnerability widely available elsewhere — was enough to remove from the market, in a matter of hours, a tool used by hundreds of millions of people. If this standard applied to the entire industry, it would, by Anthropic’s own admission, freeze most of the deployments of cutting-edge models.
Translate that into business risk: your AI provider is a provider like any other, with one difference. It can be turned off by a foreign regulator, overnight, for reasons over which you have no control.
The real risk is not technical, it is in the supply chain
The debate over the risks of AI has focused on the internal: hallucinations, data leaks, GDPR compliance and the AI Act. These are real subjects. But the Anthropic affair highlights an external danger, still rarely anticipated: the risk of supply.
Export controls, sanctions, litigation, data localization requirements, regulatory reversals — all levers that can cut off access to a model without any line of your code being affected. For a European company, the leading supplier is almost always subject to foreign law. Your continuity of service therefore depends on a chain that you do not control, and that you often have not even mapped.
The map of alternatives — and why “sovereign” ceases to be a slogan
This is where the debate on sovereign AI changes its nature. As long as everything works, relying on a sovereign provider seems like a political preference. The day a model is cut short by a foreign decision, it becomes a simple question of resilience.
Europe is not deprived. Mistral AI, founded in Paris in 2023 and valued at around 12 billion euros, is today the main supplier of models outside the United States – facing leading competitors (OpenAI, Anthropic, xAI) all American. On the infrastructure side, OVHcloud and Scaleway offer European hosting. The border remains predominantly American, and the European ecosystem is more restricted. But “more restricted” is not “useless” — it is precisely what allows you to keep an exit door when the main door closes.
Three mitigation measures to deploy immediately
1. Abstract the supplier. Build your stack to change models without rewriting everything: an abstraction layer between your applications and the API, portable prompts, tests that replay on several models. The model should be an interchangeable consumable, not a marriage.
2. Keep a hot alternative. A second supplier already integrated and tested, ready to take over in a few hours – ideally European, so as not to depend on the same rights as the first. A shift that has never been repeated is not a shift, it is a hope.
3. Map the exposure. What processes stop if a certain model disappears tomorrow morning? What turnover, what production depends on it? If the answer is “we don’t know”, that’s already the problem. AI risk must be integrated into the risk matrix in the same way as a cloud outage or the failure of a key supplier in the value chain.
In the end
AI is infrastructure, and infrastructure is designed with redundancies — not a single point of failure in another country’s capital. The question for 2026 is no longer just “which model is best”, or even “where does it run and at what cost” — it’s “who can turn it off, and how quickly can I switch over”.
Seen from this angle, the European bet on sovereign AI looks less like protectionism than basic prudence. And prudence, in an infrastructure, is not a luxury. It’s the job.
