The first time a software vendor disappeared on me, I learned a lesson it has taken the AI industry about fifteen years to catch up with: the platform you bet on can be taken away from you, and the people taking it away will not call first.

I thought about that today for the first time in years.

Last night, at 5:21pm Eastern Time, the US Commerce Department emailed Anthropic an export control directive ordering them to immediately block every foreign national on earth — including foreign Anthropic staff — from using Mythos 5 and Fable 5. Fable 5 had been publicly available for three days. Because you can’t cleanly partition a frontier-model API by passport, the only way Anthropic could comply was to disable the models for everyone. Americans included. Developers in San Francisco, London, Toronto and Peterborough all got the same screen on Friday evening: temporarily unavailable. An export control targeting foreigners had, within minutes, locked out the country that issued it.

This matters because Mythos isn’t a toy. It’s the model Anthropic have been carefully restricting to about fifty large organisations — Apple, Google, Microsoft, JPMorgan, CrowdStrike, the Linux Foundation, that level — through a programme called Project Glasswing, because in their own description back in April it had been finding high-severity vulnerabilities in every major operating system and browser it touched. Mozilla alone has used it to close hundreds of bugs. The US government has just told the world’s largest software companies that the thing they’ve been quietly using to harden their codebases is, as of Friday evening, unavailable.

The official reason is national security. The actual reason, reading between the lines of Anthropic’s unusually pointed response, appears to be that some other company managed to “jailbreak” Mythos into doing the thing it does anyway — finding flaws in software. Anthropic, complying with the order while disagreeing with it openly on their company website, noted drily that the capability in question is freely available in OpenAI’s GPT-5.5 and any number of public models, and that the defenders who keep production systems running do exactly this work every day of their lives.

So which is it: politically-motivated retaliation against a lab that’s been at war with the Pentagon all year, or a lab that spent three years lobbying for government control of frontier AI and is now surprised the government has it?

It’s both. Obviously it’s both. Anyone telling you it’s only one of those things is selling you something. The administration has been visibly furious with Anthropic since they refused to allow Claude to be used for autonomous weapons and domestic surveillance, the Pentagon stuck them on a supply-chain blacklist in retaliation, David Sacks has spent eighteen months calling them woke regulatory-capture merchants, and the IPO filing went in last month — the timing is not subtle. And Anthropic are the company that has spent the last three years arguing, sometimes loudly, in front of Congress, that frontier AI is dangerous enough that the government should have the authority to block its deployment. When the government finally exercises that authority, somewhat ham-fistedly, on you specifically, the irony is roughly the size of Belgium. You don’t get to call for a fire brigade and then complain about the water damage.

Right. Enough geopolitics. Here is what I actually care about, and what I think every CIO and CAIO I work with should care about by Monday.

Your enterprise AI stack is now dual-use technology. That is the only sentence in this blog post that matters. Internalise it. The most capable AI models are no longer in the same regulatory universe as Salesforce or Microsoft 365. They are in the same regulatory universe as cryptography and missile components. Your procurement team has not noticed this yet. Your legal team has not noticed this yet. Your continuity planning certainly has not noticed this yet.

What that means in practice is — and forgive me, because I have had this conversation roughly forty times in the last six hours with various people who should already know — find out exactly which of your production workflows fall over if Claude dies tomorrow morning. Not in a fortnight. Today. Most of the enterprises I work with cannot give me an answer in under two weeks, and that is itself the answer. If you don’t know, you’re exposed. If you do know, you can start triaging.

Then: stop building anything important on a single named model from a single named vendor. I know your engineering team has just done a beautiful integration directly against the Claude API. I know switching providers sounds like next year’s problem. It isn’t. It’s this quarter’s problem. Stick a routing layer in. Make the model swappable. Test the swap. Test it again next month. The fact that the rest of Anthropic’s stack — Opus, Sonnet, Haiku — stayed up through this whole thing is the cheerful part of the story, but the bit that broke is precisely the bit your most ambitious AI projects tend to be built on. That is not a coincidence. It will keep happening.

Sovereignty. Here is where, if you operate in the UK or anywhere in Europe, you need to take five minutes with your board. The answer to the question “can a foreign government order our AI vendor to turn us off, with no notice, no judicial review, and no appeal” is now, as of last night, definitively yes. That isn’t an abstract policy question any more. Add it to your risk register. Decide what you’re going to do about it. Mistral exists. UK AISI exists. Sovereign cloud regions exist. Whether you actually need them for your specific workloads is a separate question from whether the question is now worth asking. It is.

Take a breath, then, but don’t take it for too long. Mythos and Fable will almost certainly be restored within a fortnight — Anthropic are too important and the order looks too thin to survive contact with their lawyers. The frontier US models will remain, by a wide margin, the most capable available, and abandoning them would be a category error. The point of this blog post is not that you should stop using American frontier AI. The point is that you should stop building as if it cannot be taken away from you on a Friday evening with three minutes’ notice, because that is now a thing that demonstrably happens.

The future, which is the bit I’m supposed to write about and which is also the bit I’m least sure of: I think we’ve just watched the moment frontier AI stopped being software and became infrastructure. Not infrastructure in the slick venture-capital sense — infrastructure in the boring, regulated, geopolitically-contested sense. Submarine cables. Nuclear fuel. GPS. The companies that come through the next eighteen months in good shape will not be the ones with the cleverest agents or the most exotic prompting strategies. They’ll be the ones whose AI architecture quietly assumed, from about Saturday morning onwards, that the model their team had decided was indispensable might, with three minutes’ notice, become indispensable to someone else as well.

It is a lesson the rest of enterprise IT learned the hard way fifteen years ago. AI is about to take its turn.