A government can switch off an AI. And tomorrow, it could switch off your cybersecurity.

Résumez ou partagez cet article :

 

Friday 12 June 2026, 5:21 pm New York time. Anthropic receives a letter from the US Department of Commerce. A few hours later, Fable 5 and Mythos 5, its two most advanced AI models for cybersecurity, are disabled. Everywhere. For everyone.

What happened

The US administration invoked its export control powers and national security to demand that access to these two models be suspended for all foreign nationals, whether or not they are in the United States, including Anthropic’s own non-American employees. Technically unable to filter its users by nationality on shared cloud infrastructure, the company had to cut off access for its entire global customer base, Americans included. Amazon Web Services, which hosts these models, confirmed it had revoked access across all its regions.

The stated reason: a method for bypassing Fable 5’s safeguards (a “jailbreak”), reported to the government by a third-party company. Anthropic strongly contests the measure. According to the company, the vulnerability demonstrated would only give access to minor, already known flaws, reproducible with other public models, including OpenAI’s GPT-5.5. In its statement, Anthropic warns that if this standard were applied across the industry, it would paralyse any new deployment of a frontier model, whatever the provider.

Yet Mythos had been designed with caution: unveiled in April 2026, it was initially restricted to around fifty partner organisations under the Glasswing programme, deemed too sensitive for open release. Fable 5, launched on 9 June, was a deliberately restricted consumer version that automatically redirected sensitive requests (cybersecurity, biology, chemistry) to a less advanced model. Three days after its launch, it was switched off.

Ultimately, it matters little who is right in the technical dispute. What is striking is the mechanism: a single letter, received on a Friday evening, was enough to cut off access to a critical tool for hundreds of organisations worldwide, without notice or immediate recourse.

A precedent, not an isolated accident

Until now, US export controls mainly targeted hardware, such as chips and computing servers, bound for countries like China. This is the first time such a measure has directly hit access to a large language model already commercially deployed, and by extension all of its international customers.

In the United Kingdom, several politicians, including former ministers, pointed out that British researchers, businesses and hospitals were using the now-disabled model. In the Netherlands and in France, politicians linked the episode to the question of digital sovereignty. The most accurate analysis is probably not “should we abandon American providers?”, which would be a caricature, but rather: which critical uses can withstand an interruption, restriction or unilateral change of access decided by a foreign administration?

What if the next letter concerned your EDR?

That is the question we should be asking. Today, it is a language model used for vulnerability research. But France and Europe now rely heavily on American solutions for most of their IT security: EDR, SOC, network filtering, cloud. These solutions are effective, mature and often excellent. But they all rest on the same mechanism of dependency: one player, one jurisdiction, one administrative decision is enough to cut off access.

This is not about hidden sabotage or a “kill switch” concealed in a product. The mechanism is more mundane, and that is precisely what makes it instructive: all it takes is for the jurisdiction under which a tool is designed and operated to have the power to ban its use for a category of users. No malice required, just a regulatory framework applied unilaterally, worldwide, overnight.

Building independence by choice, not by constraint

Yet there are solid French and European alternatives, in cybersecurity as in other critical building blocks. Choosing them is neither retreat nor naive protectionism: it is a matter of operational resilience. Not every organisation is meant to rebuild everything in-house, and most everyday uses (assisted writing, customer support, document classification) are handled perfectly well by less advanced or more diversified tools. But for truly critical functions, the question deserves to be asked calmly, rather than discovered on a Friday evening under the pressure of a letter nobody saw coming.

Better to build this independence calmly, in advance, than to discover one day that you need it under duress.