Last week a government reached in and switched off the most capable AI model ever released to the public. Days after it launched.

I don't think we've quite registered how strange that is. A private company built a thing, put it out into the world, and then a phone call from Washington made it disappear. Not because it broke. Not because of a privacy scandal or a lawsuit. Because the US government decided it was a matter of national security.

The model was Anthropic's Fable 5. If you missed the story, you'd be forgiven. It moved fast and most of the mainstream coverage was buried under everything else going on. But I think this is one of those moments we'll look back on as a line in the sand. So let me walk through what happened, and then the bit that actually interests me: what it tells us about who really controls this technology, and whether "the US is weaponising AI" is the right way to read it.

What actually happened

Fable 5 launched on 9 June. I wrote about it here when it landed. It's what Anthropic calls a Mythos-class model, the top of their range, sitting above Haiku, Sonnet and Opus. The short version: it's a genuine step up. It can run on its own for hours, sometimes days, working through a problem without someone holding its hand. People who tested it early were building things in a single prompt that would normally take a small team a week.

Four or five days later, it was gone.

On around 13 June the US Commerce Department, via Secretary Howard Lutnick, sent Anthropic a directive under national-security export-control powers. The instruction was that Fable 5 and its bigger sibling Mythos 5 could not be made available to any foreign national. Not just people outside America. Any foreign national, anywhere, including foreign nationals working inside the United States, and including Anthropic's own non-US staff.

Think about how you'd actually comply with that. You can't reliably sort every user of a public AI model by passport. So Anthropic did the only thing the order really left open to them. They pulled both models for everyone. British users, American users, the lot. Switched off.

There's a detail here that matters. The US has used export controls on AI before, but always on the chips, the physical hardware that trains and runs these models. This is the first time the controls have been pointed at the model itself. The software. The weights. That's new, and it's a bigger deal than it sounds.

The reason given, and the row about it

So why? The government said it had become aware of a way to "jailbreak" Fable 5, to get around its safety guardrails.

Anthropic's account is rather less dramatic. They say the technique they were shown amounted to pointing the model at a specific codebase and asking it to find software flaws, which surfaced a handful of already-known, minor vulnerabilities. Their position is that this was a narrow trick, not a master key that unlocks everything the model can do. A senior Trump administration official, talking to Fox Business, framed it differently and put it down to Anthropic's "recklessness."

Here's the part I keep coming back to. Anthropic pointed out that the same jailbreak would almost certainly work on other public models, including OpenAI's GPT-5.5, which is sitting there fully available and under no such restriction. And they argued that if "a narrow potential jailbreak" is the bar for recalling a commercial product used by hundreds of millions of people, you've just made it impossible for anyone to ship a frontier model ever again.

You don't have to take Anthropic's side to notice the inconsistency. If the threat is real and general, why is only one company's model affected? If it's narrow and specific, why pull the entire thing for the whole planet? Both can't be true at once.

Is this the US weaponising AI? Let's be precise

The instinct, reading this, is to reach for the word "weaponising." Government seizes control of dangerous technology, points it at national-security ends. It's a tidy story.

I don't think it's quite the right word, and I'd rather be accurate than dramatic.

Weaponising would mean taking the model and turning it into an instrument of attack. That's not what this is, at least not on the visible evidence. What this actually is, is the US treating a frontier AI model the way it has long treated other dual-use technologies. Encryption. Certain chips. Nuclear know-how. Things that are useful and dangerous at the same time, and that the state decides it gets to control who can have.

In other words, AI just got reclassified. It used to be a product. As of last week, in the eyes of the US government, a top-tier model is a strategic asset. Something to be export-controlled, restricted by nationality, switched on and off according to who the administration of the day decides should have access.

That's the real shift, and honestly I think it's more significant than "weaponising" would have been. Weaponising one model is a single event. Redefining the entire category of frontier AI as a controlled export is a permanent change to the rules everyone now operates under.

Who actually controls this, then

This is where it gets uncomfortable, and where it stops being an American story.

Almost all the genuinely capable AI models in the world come from a small number of US companies. They run in US-controlled data centres. The most powerful versions aren't things you download and own. They're services you rent, accessed over the internet, hosted on infrastructure that ultimately answers to one government.

What last week proved, beyond any doubt, is that this infrastructure has a kill switch. And the hand on the switch is in Washington, not London, not Brussels, not anywhere else.

For a few days, the best AI on earth was unavailable to a British business not because of anything that business did, not because of any British law, but because of a dispute between an American company and the American government. We were simply downstream of a decision we had no part in and no vote on.

I'm not saying the decision was wrong. There may well be classified context none of us can see. But the dependency is the point. When the tool you build on can be taken away by a foreign state with no notice and no appeal, you don't really control your own supply chain. You're a tenant, and you've just watched the landlord change the locks to make a point.

What a government like ours should actually do

I'll resist the urge to be alarmist, because the honest answer is boring and that's usually the sign it's right.

No, the UK cannot match American or Chinese spending on frontier models. Pretending otherwise is a fantasy and chasing it would waste billions. But there's a real lesson here, and it's about resilience rather than rivalry.

If a single foreign government can switch off the AI your economy is quietly starting to run on, then betting everything on access to that one source is a strategic risk, not just a commercial convenience. That argues for a few sensible things. Keeping genuinely open models, the ones you can actually host yourself, as a viable fallback rather than a poor cousin. Some sovereign compute, so the country isn't entirely dependent on data centres it doesn't control. And treating "which models can our critical services rely on, and who can pull them" as an infrastructure-security question, the same way we'd think about energy or undersea cables.

None of that is about beating America. It's about not being left helpless the next time someone in Washington decides to flip a switch. Because there will be a next time. Last week established that the switch exists and that they're willing to use it.

Where this leaves us

Fable 5 might come back. The directive might get walked back, narrowed, or quietly resolved. By the time you read this the specifics may already have moved on.

But the precedent doesn't move on. We now know, for certain, that the most powerful AI in the world is something a government will treat as a controlled strategic asset, restrict by nationality, and shut off when it decides to. The technology that an enormous amount of the economy is starting to depend on can be turned off by people who don't answer to most of the people depending on it.

That's worth sitting with. Not as a reason to panic, and not as a reason to stop using these tools, which are genuinely remarkable and aren't going anywhere. But as a reason to be clear-eyed about what they are and whose hand is on the switch.

I build a lot of my own work on these models. I'll keep doing it. I'm just going to do it knowing the ground can move, and making sure I'm never standing on only one bit of it.

Peter