The Trump administration's order forcing Anthropic to restrict foreign access to its most capable AI models has demonstrated, in a single decision, that the US government now holds effective veto power over who can use the world's most consequential technology, a form of leverage with no peacetime precedent in the modern economy.

Key Highlights

  • The Trump administration ordered Anthropic to block foreign access to its most advanced frontier models, Fable and Mythos, citing a supposed jailbreak that could enable hacking or weapons development.
  • The order came just months after the administration designated Anthropic a supply-chain risk in March, establishing a pattern of willingness to intervene directly in frontier AI access regardless of commercial or diplomatic costs.
  • The episode exposes a structural dependency problem for US allies and the rest of the world: America controls not only the leading AI models but also roughly 15 times more computing infrastructure than Europe, creating a concentration of AI power with no near-term counterweight.

The Trump administration's directive to Anthropic to restrict foreign access to its latest models has settled a question that policymakers and technologists had debated in theoretical terms for years: whether the United States government has the practical authority to decide who in the world can use the most powerful AI systems. It now demonstrably does.

The stated justification was a jailbreak vulnerability in the Fable model that could potentially be exploited for cyberattacks or weapons design. Anthropic complied while arguing the concern was overstated, ultimately restricting access broadly rather than only to the specific foreign group the administration had targeted. The legal basis for the order has not been made fully public, and the ban is not expected to be permanent. But the demonstration of capability is permanent regardless of how the specific episode resolves.

This follows the administration's March designation of Anthropic as a supply-chain risk, which marked the first time the US government had applied that classification to a domestic AI laboratory. Together, the two interventions establish that the administration is prepared to override commercial interests, internal company positions, and the preferences of foreign users, all of whom had no recourse, to control AI access as a tool of state power.

The geopolitical implications extend well beyond Anthropic. Roughly 80% of Anthropic's consumer use is estimated to be outside the United States. Many of its researchers and engineers are not American citizens and were affected by the foreign-access restriction directly. Freezing AI research at a frontier laboratory would be self-defeating for American commercial interests, but the administration has shown it is willing to accept that cost when it judges security stakes to be high enough.

The structural leverage this creates over US allies has no equivalent in recent decades. America holds roughly 15 times more computing infrastructure than Europe and continues to attract the majority of global AI investment. If SpaceX realises its vision of orbital data centres, that compute gap is likely to widen rather than close. Europe, which has been building contingency plans around US reliability in defence and trade, now needs to extend that planning to AI dependency, a domain where it is far more exposed and far further behind.

The historical parallels are instructive but imperfect. Post-war nuclear technology restrictions eventually gave way to bilateral cooperation with Britain once it developed independent capability. Cryptography could not ultimately be contained and went public. Frontier AI is more like nuclear than cryptography in terms of concentration and capital requirements, but shares cryptography's dual-use ambiguity. The outcome will depend on whether rivals can develop credible independent capability, and on whether America decides its economic interest in global AI sales outweighs its strategic interest in exclusivity.