America Switched Off Its Best Model. The People It Locked Out Are the People Who Built It.

On Friday evening, at 5:21pm Eastern, the US Commerce Department sent Anthropic a letter. By the time most of the country had logged off for the weekend, the two most capable AI models on the planet, Fable 5 and Mythos 5, were dark for everyone. A lot of people found out the way I did: opening their tools and being told the model no longer existed.

The directive does not ban Americans. It bans every foreign national, wherever they live and whoever they work for, including Anthropic's own non-citizen employees. Because a company cannot verify the citizenship of every user on its platform in real time, the only way to comply was to shut the models off completely. So a national security order aimed at foreigners ended up denying the product to Americans too, which tells you something about how carefully it was thought through.

I have a stake in this I should declare. I am a foreign national. My British citizenship application is in progress. As of Friday, the official position of the US government is that I am a security risk who cannot be trusted with a chatbot. That is the frame. I want to be honest that it makes me angry, and then I want to set the anger aside and look at the mechanism, because the mechanism is the part that should worry the people who signed the letter.

What actually happened

The stated reason is a jailbreak. The government believes it became aware of a method of bypassing Fable 5's safeguards. Anthropic's own statement says it reviewed a demonstration of the technique, disagrees that a narrow potential jailbreak justifies recalling a commercial model deployed to hundreds of millions of people, and notes that applying this standard across the industry would halt new model releases for every frontier provider. Reporting indicates the same exploit also works on OpenAI's GPT-5.5, which makes the idea that Fable is uniquely dangerous hard to sustain.

The timeline does the rest of the work. Fable 5 launched on a Tuesday. It was gone by Friday afternoon, having existed for roughly seventy-two hours. The letter was reportedly signed by Commerce Secretary Howard Lutnick and cited national security authorities without specifying the concern. This is the same administration whose Defense Department labelled Anthropic a "supply chain risk" earlier this year, after the company refused to make its technology available for autonomous weapons and for mass surveillance of US citizens.

Anthropic sued. A national security export order landing on a Friday night, three days after a flagship release, citing a vulnerability the company says reveals nothing new, is not a safety process. It is leverage wearing a safety costume.

The Blockbuster comparison, and why this is worse

The obvious analogy is Blockbuster turning down Netflix, certain that nobody would ever stop coming to the store on a Friday night. The shape is right. Blockbuster did not lose because someone built a better video store. It lost because the meaning of renting a film changed underneath it, and its dominance was the exact thing that made the change invisible. The store footprint was an asset right up until it was the thing killing the company.

But the comparison flatters this situation, because Blockbuster's blindness was commercial.

It bet its own capital and lost its own business. What happened on Friday is a state reaching into the store and switching off its own champion's best product. Anthropic did not fail to invest in the future. It got ordered to take the future off the shelf. And the competitor here is not sitting inside the same market waiting to be acquired. It is Mistral in France, the open-weights labs in China, whatever the Gulf sovereign funds are standing up, all in jurisdictions the directive cannot reach. Washington can forbid an American company from serving the world. It cannot forbid the world from building the substitute. The tool that feels like control is the same tool that funds the replacement.

The advantage was never homegrown

Here is the part that turns frustration into something closer to disbelief. The American lead in AI is imported. The Paulson Institute's Global AI Talent Tracker finds that US institutions employ around 59 percent of the world's elite AI researchers, and that this lead is built almost entirely on foreign-born talent. More than half of the best researchers working in America were born somewhere else. China alone produced 47 percent of the world's top-tier AI talent in 2022, against 18 percent for the United States, and 72 percent of those China-educated researchers currently work at American institutions rather than at home.

Sit the directive on top of those numbers. The government has classified foreign nationals as a security risk to be denied the output of these models, explicitly including Anthropic's own foreign-national staff, while the entire input to the American AI advantage is foreign nationals. The engineer Commerce wants locked out of Fable 5 may well have helped build it. This is not a contradiction at the margins. It is the load-bearing beam of the whole structure.

And the one moat that is genuinely hard to copy, the thing no rival can simply buy, is being the place the world's best people want to work. The talent trackers have been warning for years that mobility is falling and that the US is becoming a less attractive destination. A competitor does not need to out-engineer America. It only needs to be the country that does not treat its smartest arrivals as suspects. Every message that says "you are a foreign national, you are a risk, you are locked out" is a recruitment email for Paris, Toronto, London and Shenzhen, written on US government letterhead.

It is not really about ego

The satisfying story is that this is Trump and a circle of yes-men too convinced of American supremacy to imagine anyone catching up. That story is not entirely wrong, but it lets the machine off too easily, because it implies you could fix the problem by changing the personnel. You could not. The output would be roughly the same with calmer egos in the room, because the structure produces it.

The pen is held by the export-control apparatus, an institution whose entire function is to prevent things from spreading. Asking it to weigh global market share is like asking a quarantine officer to think about tourism. The cost of the decision lands on foreigners and allies who do not vote, while the benefit, the appearance of toughness on national security, lands at home and pays out immediately. Nobody in that loop is rewarded for protecting a talent magnet or a global install base, because those are slow, diffuse goods measured in years, and the incentives run in news cycles. The hubris is real. It is also the flavour, not the recipe.

A cybersecurity researcher put the other half of it well when he said that if you describe your product as a munition in every press release, eventually a government takes you at your word. The industry spent a year insisting these models were almost too dangerous to release. Washington heard that and reached for the instrument you reach for with munitions.

The uncomfortable part

So here is where it lands, and it is not comfortable for anyone. The threat the directive names is foreign access to a frontier model. The people it actually harms are the allies who were never the threat and the foreign-born researchers who built the thing in the first place. The capability it is trying to contain was not a breakthrough nobody else can reach, by Anthropic's own description of how Mythos was made, which means containment buys months, not safety. And the substitute is already forming in the exact places the policy cannot touch, staffed in part by the people America is busy informing they are not welcome.

The Friday-night certainty that the world will always come to the American store is the most expensive assumption in technology. It felt true at Blockbuster too, right up until the week it reversed. The difference is that Blockbuster only took down Blockbuster. This time the store has a flag on it.

If this resonated:

I'm Gail Weiner. I work with senior leaders on the human layer of AI adoption: the place where trust either holds or breaks.

Most AI deployments fail not because the technology is wrong, but because the trust contract underneath it was never built.

The Trust Architecture Diagnostic is three 90-minute sessions designed to map where that contract is fraying in your organisation, and what to do about it. Built for senior US leaders deploying AI in their teams. Direct booking. No procurement.

Book a diagnostic → info@gailweiner.com

Read more about Trust Architecture → gailweiner.com/trust-architecture