Export controls were built for missiles and semiconductors. Using them against a remote AI service creates a regime that is invisible to outsiders, unilateral in application, and impossible to challenge publicly.
The Trump administration has ordered Anthropic to cut off all users — including American customers and the company's own engineers — from its two newest AI models, citing national security. The legal instrument it used was built for missiles and semiconductors, not chatbots you access from a laptop.
On Monday, Anthropic began blocking access to Fable 5 and Mythos 5 for every user the company determined was a foreign national. The move was not a product recall or a voluntary safety shutdown. It was an export control directive — the kind of government order more commonly used to stop weapons and chip shipments at the border. Anthropic's models don't ship anywhere. You log in from a laptop. That category mismatch is the whole story.
Anthropic said on its website that the government cited "national security authorities" to justify "an export control directive" on the models. The administration has not publicly explained the legal basis. The Verge reported that the directive appears to have cited the Export Administration Regulations, a Cold War–era framework administered by the Commerce Department, though experts who spoke to The Verge said they had not seen the actual order and could not confirm its precise scope.
The order, whatever its form, represents the first time US export controls have been used to restrict access to a hosted AI model. "To my knowledge, this is the first time US export controls have been used to control access to an AI model in this way," said Hanna Dohmen, a senior research analyst at Georgetown University's Center for Security and Emerging Technology. The administration, she noted, is improvising in real time.
Export controls have historically applied to discrete shippable items — weapons, hardware, technical data, and increasingly software and source code. These are things you can copy, download, publish, or hand to someone else. The Biden administration moved to extend that framework to AI model weights, the core parameters that make a model function. The Trump administration rescinded that rule via the Bureau of Industry and Security in 2025. What it substituted was a framework that, according to outside legal analysis, leaves the government wide latitude to act — and companies wide awake at night trying to figure out what the rules actually are.
"This is an unsettled area of export control rule-making," said Andrew Reddie, a professor at UC Berkeley's Goldman School of Public Policy. The current regime gives the government "wide latitude," he told The Verge, but leaves firms uncertain about their responsibilities. "This episode makes clear the unsustainability of the existing governance regime."
That question — what exactly Anthropic is exporting — is not rhetorical. The company hosts its models on its own servers. Users interact with them through an API or a web interface. Under a traditional export control framework, the controlled item is the thing that crosses a border. Nothing crosses a border here except electrons. The order effectively redefines "export" to include a foreign national's access to a US-hosted service — a redefinition that has no clear precedent in the regulatory record.
Congress is aware of the gap. The Remote Access Security Act is moving through the Senate, designed to extend export controls explicitly to cover remote access to controlled technology via cloud services. Latham & Watkins noted in a legal analysis that the bill, if enacted, would create an affirmative compliance obligation for companies whose services involve foreign users. But the bill has not become law. The administration appears to have decided it does not need to wait.
Three explanations are circulating among people tracking the situation. The first: Anthropic's models are uniquely capable, and the administration believes their release to foreign nationals represents a genuine national security risk — a theory that, if applied consistently, would presumably also cover OpenAI's GPT-5, Google's Gemini, and Meta's Llama releases. The second: a specific jailbreak reportedly used by groups linked to China did break some meaningful safeguard, and the administration is acting on a specific technical determination it cannot disclose publicly. Anthropic has disputed this, saying the jailbreak "did not allow users to circumvent all of the company's safeguards." The third: Anthropic is being singled out over its testy public relationship with the Trump administration, which has included public disagreements over AI safety policy and regulatory philosophy.
No outside expert, journalist, or rival company can evaluate which theory is correct, because the government's legal theory has not been made public. That opacity is not incidental. It is the governance gap.
The immediate effect is on Anthropic and its users. The broader effect is on every frontier AI lab operating in the United States. If the administration can compel a US company to restrict access to its own models — including for domestic users and the company's own employees — on the basis of undisclosed national security authorities, under a legal theory no outside expert can evaluate, then every other frontier lab is one undisclosed directive away from the same situation. The precedent does not require the government to be right. It only requires the government to act.
Foreign governments and companies are already drawing conclusions. The incident is being cited in strategic conversations in Europe, the Gulf, and Southeast Asia as evidence that the United States cannot be trusted as a reliable frontier AI supplier — that the same administration which built the world's most advanced AI ecosystem can cut off access to that ecosystem on demand, for reasons it will not disclose. Whether that characterization is fair is beside the point. It is the argument being made, and the administration has handed its advocates a structural claim they could not have made six months ago.
Georgetown's Dohmen called the legal landscape "an open question." Berkeley's Reddie called it "unsustainable." Both assessments are correct. The administration has discovered that Cold War–era export law can reach into a US company's servers and lock out users — including domestic ones — without public justification, without a challenge mechanism, and without anyone outside the room being able to say with confidence what authority was actually invoked. That is the governance gap. It is not specific to Anthropic. It is specific to the nature of the instrument the government chose to use.
The models may come back online. The legal question will not go away.