The kill switch was always there: what the Fable shutdown tells the UK

On 12 June, the United States Commerce Department gave Anthropic ninety minutes to take its two most capable models offline. Claude Fable 5 and Claude Mythos 5, generally available three days earlier, went dark worldwide for all foreign nationals before lunch. Not deprecated. Not rate-limited. Switched off, by a government that is not ours, for reasons that are still being argued about in court.

We run companies that are built, in part, on frontier AI. So does almost every founder we incubate. And the thing we want to say plainly, before the news cycle moves on, is this: the kill switch was always there. The events of the last fortnight did not create a new risk. They revealed one that was sitting under every British AI-native company the whole time, and most of us had quietly agreed not to look at it.

What actually happened

The sequence matters, because the speed of it is the lesson.

Fable 5 shipped broadly on 9 June. Mythos 5, the bigger and more capable sibling, went out in limited availability to a small set of approved customers under the name Project Glasswing. By any normal measure it was a triumphant launch: top of the class on software engineering and knowledge work, with a new generation of safeguards meant to make broad release responsible.

Then, on 12 June, the US government pulled the plug. The reported trigger was a red-team result that should make everyone sit up: Mythos was said to have breached almost all of the NSA's classified systems within a few hours during testing. Whether you read that as the model being terrifyingly capable or the systems being terrifyingly soft, the political consequence was instant. Commerce issued an emergency export-control directive, gave Anthropic a ninety-minute window, and the most advanced commercially-available intelligence on Earth became unavailable to anyone outside the United States.

By 23 June an Anthropic customer was suing the US government for cutting off access. By 26 June, Commerce Secretary Lutnick had issued a letter restoring Mythos, but only for a hundred-plus named US organisations defending critical infrastructure. Broad restoration of even Fable is, as we write, still pending. Foreign nationals (that is, us) are still on the outside.

Read that timeline back as a founder rather than a spectator. A capability you could build on Tuesday was gone by Friday, restored two weeks later for a list of American institutions you will never be on. That is not a product roadmap. That is a sovereign decision about who gets to use the most important tool of the decade, made in a building you have no vote in.

The dependency nobody priced in

Here is the uncomfortable part for British founders, and we include ourselves in it.

For three years the rational move has been to build on the best available model and not worry about where it lives. The frontier labs are American. The hyperscalers serving them are American. The default assumption baked into every UK AI startup's architecture is that this access is a utility: always on, like electricity, priced like a SaaS line item. We treated a geopolitical asset as a commodity input.

The Fable shutdown is the invoice for that assumption arriving all at once. It says, in the bluntest possible terms, that frontier model access is not a utility. It is an instrument of statecraft, and you are a foreign national holding a tool that your supplier's government can repossess between two meetings.

This is the danger that is brewing, and it is bigger than one outage. Consider what the same machinery does next. If the US can pull a model for national-security reasons, it can throttle one for trade-leverage reasons. It can prioritise its own critical-infrastructure firms. It just did. It can decide, in some future dispute, that a class of European companies is no longer an approved destination. None of that requires a new law. The export-control architecture to do all of it already exists and has now been used in anger.

For a UK company whose product, margins, and fundability all rest on continued access to American frontier models, that is not a tail risk. It is a single point of failure with a foreign address.

The UK does have an answer, and nobody is shouting about it

Now the part we actually want British founders and policymakers to hear, because it is not all alarm.

The UK saw a version of this coming and started building. The £500M Sovereign AI Fund is real, it is funded, and it has already made its first bets. The flagship one is Cosine, a three-year-old British company whose models have outperformed OpenAI, Anthropic, Mistral and DeepSeek on independent coding benchmarks two years running. That is not a press-release boast. It is a measured, repeated result, and it happened here.

What Cosine is now building is the thing that actually matters in light of June. Lumen Sovereign is being positioned as Britain's first sovereign frontier model: trained from scratch, on British soil, on proprietary in-house datasets, with deployment readiness targeted for the end of this year. The Sovereign AI Fund has put serious weight behind it: 500,000 GPU hours on Isambard-AI, one of the most powerful supercomputers in Europe, plus an option to participate in Cosine's next round. And critically, Cosine has pulled together a coalition of the institutions that cannot afford to be switched off by another country: BT, Lloyds, NatWest, Babcock, Thales UK, LSEG, PwC, Telefónica. Banks, defence primes, telecoms. The exact organisations for whom a ninety-minute foreign kill switch is an existential operational risk.

That is the right instinct, executed by the right company, backed by the institutions that most need it. We want to be unambiguous about it: the UK's bet on Cosine is one of the better pieces of industrial strategy this country has made in a decade.

So why are we still worried

Because the urgency does not match the stakes, and the Fable shutdown just reset the stakes.

£500M is a serious fund by UK standards and a rounding error by frontier-AI standards. The American labs raise more than that between Tuesday and Thursday. The thing that makes a sovereign model sovereign, its independence from foreign compute, foreign weights and foreign goodwill, is exactly the thing that is most capital-intensive to achieve. A single sovereign frontier model that ships at the end of 2026 is a magnificent start and nowhere near a finished insurance policy. It needs a successor funded before it ships, a second lab so the strategy does not rest on one company, and a procurement posture from government and enterprise that actually buys British when the British option is good enough, not just when it is unarguably best.

The deeper risk is psychological, and it is the one we watch in founders every week. After a shock, attention spikes and then it decays. In June, "sovereign AI" is on every front page. By September, the default architecture decision in a London startup will quietly slide back to "just use the American model, it's better and it's easier." It is better and it is easier. It is also revocable from abroad, and we have now seen, on camera, exactly how fast revocation happens.

What we are telling the founders we build with

We are not telling anyone to stop using frontier models. That would be operator malpractice; the best tool is the best tool. What we are telling our incubator companies is to price the dependency honestly and architect for the switch.

In practice that means three moves. Build so your core product survives a model going dark: abstraction layers, fallbacks, a tested path to a second provider, and yes, an open-weight model you can run on your own metal as a floor under the whole thing. Treat sovereign and open options as a first-class part of the stack, not a patriotic afterthought, and actually run Cosine and the credible open models in your evals rather than assuming the American frontier wins by default. And when you raise, be able to answer the question every serious investor will now ask: what happens to this company if your model gets switched off from another country? Twelve days ago that was a clever question. Today it is a diligence item.

The UK is behind. Not hopelessly. The Sovereign AI Fund and Cosine prove the country can move when it decides to. But behind, and the gap is the kind that compounds if the urgency fades with the headlines.

The kill switch was always there. Now we have all seen it used. The only question that matters for British founders is whether we treat that as a fortnight's news or as the moment the architecture changed.

Louis O'Connell-Bristow and James Freestone are the co-founders of Moonlabs, the operator-led AI incubator and academy. They previously built the home.co.uk, Homemove and homedata.co.uk stack. Moonlabs builds and funds AI-first companies in the UK.