Trump orders Anthropic to take Mythos 5 and Claude Fable 5 offline, allies brace

US President Donald Trump’s administration ordered Anthropic to cut off foreign access to its advanced AI models Mythos 5 and Claude Fable 5, and told the company to take the models completely offline to comply. The order applies to “all foreign nationals in and outside the US” and was issued last week, immediately jolting Europe, where US-developed AI is deeply embedded.

The practical effect is blunt: Anthropic previously granted 200 institutions across 15 countries access to its frontier model, Claude Mythos Preview, to test for vulnerabilities. The two public versions of the model, Mythos 5 and Fable 5, were due to be released in early June. Anthropic said the US government did not provide a reason, but that it was operating on its “understanding” that the administration believed it had become aware of a method of “jailbreaking” Fable 5.

Why this matters is not only about one company. It is about who gets to run the most advanced parts of the AI stack, and under what conditions. The Trump administration’s move is described as unprecedented for the AI industry, but it lands in a broader pattern: over the past 18 months, Trump has launched a global trade war, threatened to annex Greenland, warned about withdrawing from the NATO alliance, and even threatened to stop supplying weapons to Ukraine unless European allies helped reopen the Strait of Hormuz, which Iran effectively shuttered after the US and Israel launched their war on Iran on February 28.

So when the US targets advanced technology access, allies do not just model technical risk. They model geopolitical risk. The source reports that the ban applies equally to allied countries that have intelligence-sharing and mutual defence pacts with Washington, which is precisely what has alarmed policymakers. Dex Hunter-Torricke, president of the Center for Tomorrow, told Al Jazeera that US allies are now realizing they are “far too vulnerable now to the US techno-industrial complex.” French President Emmanuel Macron called the order a “wake-up call” on AI dangers, but said the limits were a “bad thing,” and criticized how the reaction can become “strictly nationalist.”

At the Group of Seven meeting, Macron also pushed for countries to work together on AI, warning against “non-cooperation between democracies.” Thomas Regnier, a European Commission spokesperson for tech sovereignty, echoed that security concerns are a “shared challenge, not one confined to a single jurisdiction or country,” and added that solutions should not be “discriminatory against partners.” In other words, Europe’s challenge is simultaneously technological and political. It wants stronger defenses against threats like model jailbreaking, but it does not want a system where access is conditional on Washington’s changing comfort level.

The G7 discussion centered on a potential “trusted partner” scheme for access to the most advanced AI models, though few details have been disclosed. This idea is showing up because the US is not the only one using tiers. The source notes that the Trump administration briefly introduced a “small yard, high fence” approach to semiconductor chips in early 2025 during the final weeks of Joe Biden’s presidency. That model was intended to keep the most advanced US technology away from China and Russia, while it also created “disquiet” among allies: it did not restrict allied access to semiconductors, but it restricted what allies could do with those chips, including commercial use or trade dealings with China. The scheme was scrapped in May 2025, and the administration later cleared Nvidia H200 for sale to a limited number of Chinese firms.

That background helps explain why Anthropic’s ban is accelerating the self-reliance conversation, even among governments that share security ties with the US. Canadian Prime Minister Mark Carney warned that the situation with Mythos and Fable could happen due to overreliance and said, “Nobody has done anything wrong,” but that accepting it without building and diversifying would be a mistake. Bruno Retailleau, a former French minister and 2027 presidential election candidate, said the ban should be a “wake-up call” that a nation depending on others for technology can be “unplugged overnight,” and he argued AI should be treated like nuclear power as part of sovereignty. Tom Tugendhat, a former security minister under Liz Truss and Rishi Sunak, framed it as a shift where sovereignty becomes “more about code than cannons.”

Meanwhile, European governments are also looking at how to reduce friction and dependencies in the broader defense and intelligence ecosystem. The source says German media reported in April that Germany’s armed forces would not award contracts to Palantir, Peter Thiel’s big data analytics company, citing concerns about granting private industry access to national data systems. It also says Germany and France’s domestic spy agencies have partnered with European companies over Palantir to avoid becoming too dependent on US technology. And as European officials discuss AI sovereignty, the ban has drawn attention to Paris-based Mistral as the “EU’s only major homegrown frontier-model competitor,” according to Jerzewski, head of the Taiwan office of the European Values Center for Security Policy.

The second-order implication for leaders on both sides is uncomfortable: the architecture of trust in AI is becoming a national security asset, not just a product feature. If advanced models can be removed from allied users with little warning and no detailed explanation, boards and executives will need to treat model access as a continuity risk, not a procurement line item. For governments, that likely means more investment in domestic capabilities and new access frameworks like “trusted partner” schemes. For companies, it means preparing for policy-driven access shifts, because the source makes clear this kind of disruption is now explicitly on the table, even for allies.