Friday 12 June, 17:21 in Washington. The US government orders Anthropic to block access to its two most powerful models for all non-US citizens. But an API doesn’t check passports. A few hours later, Claude Fable 5 and Mythos 5 are taken offline for everyone. This is the first time a government has shut down an artificial intelligence system that has already been made available to the public.

From OpenAI to Anthropic

In 2021, Dario and Daniela Amodei left OpenAI, arguing that the pace at which the company was developing its models was incompatible with safety. They immediately founded Anthropic, now the parent company of Claude and OpenAI’s main rival, with safety as its founding mission.

The Clash with the Pentagon

In 2025, the Amodei family’s company signed a $200 million deal with the Pentagon, but on two conditions: no mass surveillance of American citizens, and no lethal autonomous weapons. Claude was thus integrated into the Pentagon’s systems and, according to reports in the Wall Street Journal and the Washington Post, was used in the operation that led to Maduro’s abduction and, in February, in targeting attacks against Iran, identifying over a thousand targets within the first 24 hours.

However, when the Pentagon tried to go beyond the agreed limits, Anthropic did not give in. This led to a clash with Washington, which went so far as to threaten the company with being designated a ‘supply chain risk’ – a label usually reserved for foreign companies deemed hostile. It was against this already tense backdrop that, months later, Fable 5 and Mythos 5 entered the scene.

What are Fable and Mythos?

On 9 June, Anthropic released Fable 5, which, according to the company, is the most capable model ever deployed on a large scale. Its capabilities centre primarily on complex reasoning, coding and agentic behaviour – that is, the ability to carry out complex tasks autonomously, without constant supervision.

Fable is regarded as a ‘tamed’ version of a far more powerful model, Mythos, which has, however, never been made public: since April 2026, its use has in fact been restricted to around fifty companies and government bodies grouped together under Project Glasswing. The reason for this caution is that Mythos possesses cyber-offensive capabilities never seen before, to the extent that, in internal development tests, Anthropic observed it identifying thousands of vulnerabilities in just a few hours, as confirmed by the UK AI Security Institute.

Just three days before the launch of Fable on 12 June, the US Department of Commerce decided to ban the use of both models on national security grounds.

The pretext

However, the US Department of Commerce’s ban was not limited to users of Fable and Mythos but was extended to anyone who is not a US citizen, including Anthropic employees without a US passport. It all began with a letter that Commerce Secretary Howard Lutnick sent to Dario Amodei on 12 June; since then, Fable 5 and Mythos have been subject to immediate export controls. These models are primarily used via APIs – interfaces through which software connects directly to the model and which function as a gateway that responds to anyone who queries it. Unable to distinguish in real time between US and non-US users, Anthropic has chosen to shut down both models for everyone.

The letter contained no technical details regarding the alleged threat to national security, and an account that circulated in the hours that followed suggested that a rival company had demonstrated to the White House a ‘jailbreak’ – a technique that forces the model to bypass its own security restrictions – on Fable 5. Anthropic responded curtly, arguing that this was a minor and already known vulnerability, which could also be exploited on GPT-5.5, a model that remains accessible.

Furthermore, according to Axios sources, the administration had already attempted to block the launch of Fable 5 even before it was released. The alleged jailbreak may therefore have been more of an opportunity seized on the spot than the actual reason for the block. For those living thousands of kilometres away from Washington, one question inevitably springs to mind: if the ban doesn’t even spare the company that created them, what does that mean for those who simply use these models?

It’s not a problem with the users

The fact that some developers have found their API keys disabled is merely a symptom. The real problem is that an entire continent has discovered that it is leasing its digital infrastructure from a government that can terminate the contract without notice and without explanation.

And this didn’t happen with a company that was in line with the US government – quite the opposite: just a few months ago, it showed itself willing to take its own government to court rather than give in. But that wasn’t enough. A single letter was all it took.

The logic behind the cancellation of Fable 5 extends beyond cutting-edge LLMs. Underlying these models is the infrastructure that underpins everyday life in Europe: government departments, hospitals, banks and universities all run on AWS, Microsoft Azure and Google Cloud. Three American companies control over 70 per cent of the European cloud market. And there’s no need for a special letter here, because the CLOUD Act already stipulates this by law: an American provider must comply with requests from Washington wherever the data is stored, and no ‘European Sovereign Cloud’ from Amazon changes the nationality of its owner.

Farrell and Newman have called it ‘weaponised interdependence’: whoever controls the network nodes turns access into a weapon. In 2025, Microsoft suspended services to users subject to sanctions. The Anthropic case does not open a new door. It merely shows just how wide the one that was already open really is.

Is technological sovereignty feasible?

In an almost comical coincidence, nine days before Lutnick’s letter, the Commission had presented its most ambitious package on technological sovereignty. But political will alone is not enough: the treaties prevent the Union from acting as a single power.

Research and technological development are a parallel shared competence: Brussels can provide funding and coordination, but it cannot centralise decision-making or bypass the twenty-seven Member States. The estimated 300 billion for self-sufficiency cannot be mobilised in the same way as in Washington or Beijing – through a top-down decision – but must be negotiated on a state-by-state basis, capital market by capital market.

Meanwhile, the assets do exist and are by no means insignificant: ASML and its monopoly on machinery for the most advanced chips, the JUPITER supercomputer now operational in Jülich, and above all Mistral, the continent’s only cutting-edge laboratory. With a fraction of the capital raised by their American rivals, their models are hot on the heels of Claude and GPT less than a year later, and ASML’s entry as the largest shareholder in 2026 demonstrates that European industry has stopped leaving its champions to Silicon Valley.

But these resources are scattered, and the figures bear this out: Europe controls less than 5 per cent of the world’s cutting-edge computing power. JUPITER is fitted with 24,000 Nvidia chips, whilst xAI’s Colossus alone has over 200,000.

There is one lever available. Article 122 of the Treaty – the solidarity clause for supply crises – was repurposed in 2025 as a geopolitical tool through SAFE, the 150 billion in defence loans. The same could be done for technology. But this is the exception that proves the rule: Europe needs an emergency to centralise what the United States does as a matter of course.

The European crisis

In the short to medium term, Europe will not achieve full technological sovereignty. This is not due to a lack of engineers or billions, but because of the way it is structured. As long as research remains an area that Brussels can only fund and never direct, these three hundred billion will not be mobilised: they are negotiated, twenty-seven times over, whilst Washington makes a decision on a Friday afternoon.

Achieving digital sovereignty does not mean chasing after a frontier model, but rather restructuring the Union’s architecture.

The kill switch on 12 June, which shut down an AI model, raised a question that Europe would rather not hear: do we want sovereignty, or just the right to complain when it is taken away from us?