AI as Ammunition: How Washington Pulled the Plug on Anthropic – and What This Means for Europe

Yesterday morning we activated Claude Fable 5 and Mythos 5 – Anthropic's brand-new flagship models – for our users at Evoya. The excitement was enormous: Fable 5 is considered the most capable language model ever made publicly available, and we were proud to have it live in the Evoya Workspace so quickly. A few hours later, we had to deactivate it again. Not because of a technical problem. But because the US Department of Commerce had instructed Anthropic via directive to block access for all non-US citizens worldwide – effective immediately.

No technical incident. A political signal. And it should make everyone in Europe sit up and take notice.

What Exactly Happened

On June 12, 2026 – three days after the launch of Fable 5 and Mythos 5 – Anthropic received a directive from the US Commerce Department on a Friday evening. The content: no more access for "foreign nationals" (non-US citizens), regardless of where in the world they are located. And not just external customers: Anthropic's own employees without US citizenship were also affected by the directive.

In their official statement, Anthropic writes that they disagree with the directive. The government's justification – an alleged jailbreak technique that could be used to circumvent the model's safety mechanisms – was described as "narrow" and "non-universal." Other publicly available models could have similar vulnerabilities without anyone ever having intervened. Anthropic considers the reaction clearly disproportionate.

Nevertheless, Anthropic complied with the directive. And chose the safe path: blocking access for all users – including US citizens – to eliminate any risk of accidental disclosure. Fable 5 and Mythos 5, just celebrated as crown jewels in the model portfolio, are thus effectively taken offline for the entire international customer base.

Adding to the tension: this happens against the backdrop of an already strained relationship between Anthropic and the Trump administration. The US Department of Defense had previously reportedly classified Anthropic as a "supply chain risk" – a designation normally reserved for hostile foreign actors. Anthropic has sued over this designation.

Context for the US concerns is provided by Project Glasswing: Anthropic launched the program after its Claude Mythos model was capable of automatically scanning third-party codebases for software vulnerabilities – identifying over 10,000 critical security flaws in real-world systems in just one month, including in critical infrastructure such as energy, water, and healthcare systems. Glasswing was meant to use this capability for good: systematically finding vulnerabilities before attackers do. But the same realization made it clear what happens when such a model falls into the wrong hands. Anthropic itself warned that many providers will soon have similarly capable models – possibly without the same safety precautions. From the US perspective, that is apparently argument enough to lock down Fable 5 and Mythos 5 as quickly as possible.

Déjà-vu: We've Been Here Before

Those who know the history of technology will rub their eyes. We've seen this before – only back then the protagonist wasn't called "Fable 5" but PGP.

In 1991, Phil Zimmermann published Pretty Good Privacy, a program for strong communication encryption. What he didn't anticipate: in the US, strong cryptography was classified as munitions under the Export Administration Act – equivalent to anti-tank missiles. Export was prohibited. Zimmermann was put under pressure by US authorities for years because his code had "crossed" the national border.

"Until 1996, cryptographic tools were classified as munitions in the United States, with strict limits on the type of encryption that could be exported."

The parallels to today are striking:

The Crypto Wars ended in defeat for the controllers: in 1999, export restrictions were largely lifted because the US had realized it couldn't stop the spread of strong cryptography – and was only hurting its own industry in the process. Will this learning process happen faster with AI? At the moment, it doesn't look that way.

What This Means for Switzerland and Europe

One sentence sums it up: Access to top AI now depends on your passport, not your subscription.

This is no exaggeration. We pay for Anthropic access, have contracts, are reliable customers – and yet access can be revoked from one day to the next because a bureaucrat in Washington signs a directive. As a Swiss company, we simply have no say in the matter.

This reveals a structural problem that extends far beyond this individual case: Europe is not sovereign in the field of AI. We have no frontier models of our own that can compete with GPT-5, Gemini Ultra, or Fable 5. There are encouraging signs – Mistral from France is making remarkable progress, and various European open-source initiatives are growing – but at the absolute performance frontier, Europe is still missing. What we can do is diversify our suppliers: instead of relying on a single US provider, build a broad portfolio of US, European, open-source, and on-premise models. That is exactly what we do at Evoya.

But let's not kid ourselves: Diversification is no substitute for sovereignty. It is the best available strategy – as long as genuine digital independence for Europe remains wishful thinking.

The Unintended Consequence: China Benefits

Here lies the real irony. The US justifies export controls on national security grounds. But what happens when European and Asian users can no longer access the best US models? They switch.

DeepSeek, Qwen, Baidu Ernie – Chinese providers have made massive strides in the last 18 months. DeepSeek, for example, has impressively demonstrated with its R1 and V3 models that world-class AI no longer needs to come from Silicon Valley. These models are powerful, affordable – and available without a passport check. The US export controls, designed to prevent "dangerous AI knowledge" from falling into the wrong hands, are driving the international user base directly into the arms of Chinese alternatives. This is likely the exact opposite of Washington's stated goal – at least if you take the official justifications at face value.

As with the Crypto Wars of the 90s: technology cannot be contained within national borders. The only question is who benefits from the gap left behind.

Our Response

For Evoya, the direction hasn't changed – but this incident confirms our course. We consistently integrate models from a wide variety of sources: US providers like Anthropic, OpenAI, and Google, but also European models, open alternatives, and Swiss and EU-local deployments for situations where data protection and availability are non-negotiable. This is exactly what our Evoya Privacy Shield stands for: customer data is automatically anonymized before any model – regardless of origin – sees it.

We will activate Fable 5 and Mythos 5 as soon as Anthropic restores international access. We monitor the situation daily.

What this incident means in the long term, we will all feel: the era in which AI models were simply available to everyone – like the open internet – may be coming to an end. And that, of all things, at the hands of the country that made the open internet into global infrastructure.

History repeats itself. Then it was PGP. Today it is Fable 5. The difference: this time it's not about private communication – but about the intellectual infrastructure of the global economy. What do you think: how should Europe respond?