The order arrived at 5:21 p.m. on a Friday, stamped with a national-security authorization, and it left Anthropic with little choice. On June 12, the U.S. government directed the AI company to cut off access to its two most capable models, Fable 5 and Mythos 5, for any foreign national, anywhere in the world, including Anthropic’s own non-U.S. employees. The company complied, in what people close to it describe as the most drastic product decision in its history.
The definition of foreign national in the directive is unusually broad. It applies whether the person is inside or outside the United States, and it reaches into Anthropic’s own engineering teams, where researchers of many nationalities build and monitor the very systems now walled off from them. The practical effect, employees said, is that a large share of the company’s workforce has been locked out of its best models overnight.
The order stems from a dispute over what the government calls a jailbreak and security researchers call a routine request. Fortune has reported that the trigger was the phrase “Fix this code,” which researchers used when asking the models to repair known vulnerabilities. The government treated the request as an attempt to circumvent safety controls; the researchers say it was a request for a patch, the kind of work the security industry has done for decades.
The episode has become a defining test of how AI governance works in practice. For years, the leading labs governed their models through voluntary commitments, safety frameworks and internal review boards. The Anthropic order replaces that system with a hard export control issued under national-security authority, and it raises a question the industry has not had to answer before: what happens when the government decides a company’s own models are too dangerous for the company to share with its own people?
The consequences are already visible. Anthropic sells access to its models to enterprises around the world, and an export ban on its two flagship systems gives international customers a reason to look at alternatives from OpenAI, Google or the open-source community. The company’s international staff face the choice of relocating, changing projects or leaving, and rivals are openly courting the talent that the ban has made idle.
The politics are layered, and the Amazon connection makes them stranger. Andy Jassy, Amazon’s chief executive, has voiced safety concerns about frontier models to the administration, according to people familiar with the matter, and Amazon is both one of Anthropic’s largest investors and, through its own models, one of its competitors. Anthropic has agreed to meet with the Trump administration to resolve the dispute, but the company’s backers and its rivals are pulling in different directions, and the meeting is unlikely to settle all of it.
For the broader industry, the order is a precedent. If the government can bar foreign nationals from one company’s models, it can do the same to any lab, and the export-control regime that has governed chips for years is now being applied to software and to the people who write it. The labs that sell frontier models abroad are watching closely, because the rules being written for Anthropic will be applied to them.
The security-research community has responded with rare unity. Seventy-six cybersecurity experts signed an open letter urging the White House to reverse the ban, arguing that treating patch requests as attacks will chill the research that keeps critical infrastructure safe. Their argument is straightforward: the flaws that AI models can help fix do not disappear because researchers are barred from asking, and the vulnerabilities will be found by someone else.
Anthropic’s own position is delicate. The company built its reputation on safety, and it cannot publicly fight an order issued in the name of national security without appearing to argue that its models should be everywhere. Its choice to comply quietly and negotiate privately is the rational one, but it comes at a cost to its international business, its talent base and its standing as a company that puts safety first.
The models at the center of the dispute are, by most measures, the most capable systems Anthropic has built. Fable 5 and Mythos 5 represent the company’s frontier work, the models it has pointed to as evidence that safety and capability can advance together, and their removal from circulation removes the showcase for that argument. The company has said it will keep the models running for domestic customers and researchers, but the ban’s terms make even that operationally difficult for a workforce that spans nationalities.
The dispute has also opened a new front in the government’s AI policy. The export-control machinery that restricts advanced chips now extends to software, and the question of who may touch the most powerful systems has become a matter of national security rather than company policy. Industry lawyers said the order’s language will be studied for years as the template for how the United States controls the frontier models it helped create.
The meeting with the administration will determine how the episode ends. If the government narrows the ban to genuinely harmful uses, the industry will treat the episode as a scare and move on. If it holds the line, the question of who may touch the most powerful AI systems will become the central argument of the frontier, and Anthropic will be the company that proved it could be decided by a document stamped at 5:21 p.m. on a Friday.
