For years, technology export controls meant things that arrived in wooden crates. Advanced chips. Lithography machines with German optics and Dutch precision. Missile guidance systems. Cryptography hardware. Physical objects with heft and serial numbers, moving through ports and customs declarations and the comforting paperwork of the material world. The entire legal architecture of export control was built on the assumption that dangerous technology has mass.On 13 June 2026, the United States effectively declared that a sufficiently capable AI model belongs in the same category as all of the above. No crate required. Anthropic was directed to disable its two most capable AI models, Fable 5 and Mythos 5, for any foreign national — whether located inside or outside the United States, including the company’s own non-citizen employees. The directive came from Commerce Secretary Howard Lutnick in a letter to CEO Dario Amodei. Anthropic, finding it operationally impossible to screen hundreds of millions of users by nationality in real time, did the only mathematically coherent thing available: it disabled both models for its entire global customer base and defaulted everyone to the older Claude Opus 4.8.
This is the part where it is worth pausing, because the individual details are almost distracting from the larger shift. The United States has now extended the Export Administration Regulations — the same framework that governs whether ASML can sell a machine to a Chinese foundry — to govern whether a software model can respond to a prompt typed by someone holding the wrong passport. The model does not cross a border. The weights do not travel in a container ship. The inference happens on American servers. The access was the export.Empires have always understood that controlling access to capability is more durable than controlling territory. The British controlled ports, then shipping lanes, then energy supplies. Washington may have just discovered something more elegant: control access to reasoning itself.The US government’s technical justification is where the story becomes either farcical or chilling, depending on your disposition. Anthropic’s blog post is worth quoting directly: the company said it had received only verbal evidence of what it described as “a potential narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws.” To translate from the security research: someone demonstrated that if you ask Fable 5 to look at code and find bugs, it will find bugs. This is, to reiterate, the advertised capability of the product. It is on the website.
Anthropic reviewed the report it believes formed the basis of the directive and confirmed that the capability level displayed was widely available from other models, including OpenAI‘s GPT-5.5, and is used every day by the defenders who keep systems safe. The company’s position, stated with restraint that must have required considerable internal effort, was that if this standard were applied across the industry, it would essentially halt all new model deployments for all frontier model providers.The blog post ends with a sentence that deserves to be read slowly: “This action does not adhere to those principles.” The principles in question being that government oversight of AI should be transparent, fair, clear, and grounded in technical facts. (Read the blog)The context, as always, is doing significant structural work. Earlier this year, the Pentagon had already designated Anthropic a “supply chain risk,” barring the military and defence contractors from using its models — a designation Anthropic is contesting in federal court. The administration had also tried to get Anthropic to pause releasing Fable 5 and Mythos 5 altogether, but was unsuccessful, which directly prompted the export control letter. The Friday evening timing, arriving hours after the models went live to global acclaim, was either a coincidence or a message. The Commerce Department is not known for coincidences.There is an irony here that is almost too neat to be comfortable. Dario Amodei has been among the most vocal advocates in Silicon Valley for more export controls on AI hardware. He wrote the op-eds. He made the congressional appearances. He argued, with genuine conviction, that the United States needed to be more aggressive about keeping frontier capability out of adversarial hands. He was not wrong, exactly. He simply did not anticipate that the machinery he was helping to build would one day be aimed at his own front door.
Anthropic has by its own telling been the first frontier AI company to deploy models on classified government networks and national laboratories, and its Project Glasswing cybersecurity initiative has already expanded to about 150 organisations across more than 15 countries. This is not the profile of a company indifferent to national security. It is, rather, the profile of a company that has discovered that being a national security partner does not protect you from being treated as a national security concern.The deeper argument here is not really about Anthropic at all. The entire AI boom has been built on a chain of assumptions: a model is software, software is infinitely replicable, replicable things become commodities, commodities become democratised. What the American government has now effectively said is: no, a sufficiently capable model is not software. It is strategic infrastructure. Once that principle is accepted, frontier AI begins to resemble nuclear enrichment, or satellite imagery, or cryptography in the 1990s, or semiconductor manufacturing today. The Anthropic story is merely the first visible symptom of a much longer reclassification.For thirty years, the internet promised that geography mattered less. The most successful technology companies in history were built on that assumption. A teenager in Gurgaon could use the same software as a banker in New York. A researcher in Munich could access the same tools as a student in Manila. The Anthropic directive suggests that frontier AI may evolve in precisely the opposite direction. Geography is returning. Nationality is returning. The passport, that stubborn artefact of the nineteenth century, has entered the chat window.For every developer in Bengaluru, Berlin, or Beirut who woke up on Saturday morning to a 404 where their most capable AI model used to be, the lesson is practical and pointed. But the truly unsettling idea sits underneath the outage, the jailbreak report, and the legal wrangling about model weights and nationality checks. For the first time in the internet era, a government has asserted that asking a question can itself be an export-controlled activity. A wooden crate, at least, you can see coming.
