Anthropic shut down its two most powerful AI models on Washington's orders. Three days after launch, and for the whole world

On Friday evening, American firm Anthropic shut down its two most powerful artificial intelligence models worldwide — the publicly available Claude Fable 5 and its unrestrained enterprise version Mythos 5 — just three days after their launch. According to the company's statement, the reason is a directive from the US government, which invokes "national security" and, under export controls, ordered that all foreigners be cut off from the models, including the company's own foreign employees. Since Anthropic says it cannot sort users by passport, it had to "rapidly shut down" the models for absolutely everyone — and this happened in the very week the company is heading to the stock exchange with a valuation approaching a trillion dollars.

The order came at 17:21

Anthropic states that it received the directive on Friday at 17:21 Eastern US time (23:21 our time) and that the government "did not specifically state the reasons for its national security concerns." According to the American outlet Axios, the order came from Secretary of Commerce Howard Lutnick, who reportedly issued it after learning that one of the company's clients had managed to bypass the model's safety guardrails. The US Department of Commerce did not immediately respond to a request for comment from either AFP or the BBC. Le Monde ↗

The company disputes the decision. It claims that the discovery of one possible guardrail bypass "does not justify withdrawing a commercial model deployed for hundreds of millions of people," and calls the whole matter a "misunderstanding." It warns that if the same standard were applied across the entire industry, it "would essentially halt all new deployments" of frontier AI models. It wants to restore access "as soon as possible." Le Figaro ↗

"Fear marketing," or a real threat?

The security concern centers on a so-called jailbreak — bypassing the software guardrails that prevent misuse. Anthropic states that it reviewed a demonstration of this technique and that it revealed only "a small number of already known, minor vulnerabilities," which are moreover "fairly simple" and can also be found by other publicly available models. BBC ↗

The catch is that Anthropic itself set off the rhetoric about danger. It launched Fable 5 on June 9 as a publicly adapted version of the Mythos family, which it had already unveiled in April and itself described as "too powerful to release." When announcing it, the company voluntarily restricted access in areas such as cybersecurity and, more recently, risks of biological or chemical misuse. Part of the expert community already read this at the time as "fear marketing" and exaggerated self-portrayal. The same framing that was an advertisement in April thus became the pretext for a government kill switch in June. Le Monde ↗

The irony is compounded by the company's own CEO. Dario Amodei had only days earlier publicly called for governments to have the authority to stop risky artificial intelligence. Washington took him at his word — and stopped his own. Der Standard ↗

War with the administration

Friday's intervention is not an isolated incident. Anthropic had already clashed with the Trump administration before: in February, Secretary of Defense Pete Hegseth labeled the company a "supply chain risk" — according to the BBC, the first time an American company received a label that Washington usually reserves for companies from hostile countries. This happened after Anthropic refused to remove the guardrails preventing use of its AI for autonomous weapons and domestic surveillance. The company sued the Pentagon, and a court has so far ruled that the order cannot be enforced. Donald Trump himself has also publicly criticized Anthropic. BBC ↗

Who holds the kill switch: a European angle

For European newsrooms, the most interesting thing about the whole case is one point — it showed how easily a single American minister, using the lever of export controls, can switch off overnight a tool that companies and institutions around the world already rely on. Le Monde, for instance, noted that the French groups Orange and Crédit Agricole were also negotiating access to the Mythos model. Le Monde ↗

The outage also came amid the frenzy surrounding tech giants going public. Following its May funding round, Anthropic reportedly reached a (post-money) valuation of around $965 billion according to US media, and it is set to hit the trading floor alongside SpaceX, which debuted on June 12, and OpenAI — this trio is expected to be the year's biggest stock market events. The paradox is glaring: one of the world's most valuable startups is not entirely in control of its own product. Le Figaro ↗

Europe reacts: a tool Washington can switch off overnight

Over the weekend, the technical curiosity turned into a political issue. The pan-European channel Euronews summarized the wave of reactions, in which European politicians agreed on one thing — the case is a "wake-up call." French former interior minister and presidential candidate Bruno Retailleau wrote that "a nation that is technologically dependent on others is a nation that can be disconnected overnight," and called on Paris to strengthen domestic companies such as Mistral and OVHcloud. Euronews ↗

Others joined in. French Minister for Europe Benjamin Haddad declared that Europe "cannot settle for the position of an open market dependent on technologies designed, financed, and controlled elsewhere." Former Prime Minister Édouard Philippe compared computing infrastructure to electricity or the internet — "infrastructure whose models and computing power we do not control is infrastructure that others can switch off for us." In France, where the presidential campaign is getting underway, the case has become one of its themes, and National Rally leader Jordan Bardella also called for Mistral to be strengthened. Euronews ↗

The German tech outlet Heise pointed out that Europe and the US approach risk differently. While Washington reached for export controls as a kill switch, the EU relies on the AI Act regulation — that is, on market oversight, transparency, and documentation obligations, where a "global shutdown of a single model" does not feature in its logic at all. Oversight in Germany is to be handled by the Federal Network Agency. The Fable 5 and Mythos 5 models remained offline over the weekend, and Anthropic and the US government were exploring whether they could be restored with adjusted guardrails. Heise ↗

What hotinfo is following

• Whether and when Fable 5 and Mythos 5 will return — and whether they will be available selectively based on user nationality.
• Whether the Department of Commerce will formally justify the directive, or stick with "misunderstanding."
• Further developments in Anthropic's lawsuit against the Pentagon.
• Whether European clients, after this demonstration of the American "kill switch," will accelerate the search for sovereign AI alternatives.
• Whether the intervention will shift the timing of Anthropic's announced stock market listing.