The White House told OpenAI and Anthropic to hold new AI models back from the UK — here's what that means

If you've never heard of the AI Security Institute, that's fair enough — it's a government body, not a consumer product, and it doesn't have a press office shouting about its work. But it does something that affects you directly: before a company like OpenAI or Anthropic releases a genuinely powerful new AI model to the public, this UK institute — alongside its US counterpart — is one of the places that gets to check it first for dangerous capabilities. This week, that arrangement changed, and it's worth understanding exactly how.

Set up by the UK government in 2023 (it was originally called the AI Safety Institute, renamed in 2025), AISI's job is to get access to powerful AI models before they go public and test them for things a company might not want to advertise: can this model help someone build a cyberweapon, assist with a biological threat, or act in ways nobody can predict or switch off. It isn't a regulator — it can't block a model from being released — but its findings feed straight into government policy, and the companies have, until now, generally cooperated voluntarily because it's good for public trust.

AISI is also, by the numbers, a serious operation: £66m a year in government funding, priority access to more than £1.5bn of UK supercomputing capacity, and over 100 technical staff, according to its own website. That matters for what follows, because the story here isn't that the UK's tester is out of its depth.

Politico's Sophia Cai and Joseph Bambridge reported on 24 September 2026 that the White House's Office of the National Cyber Director asked OpenAI and Anthropic not to share their newest models with AISI until American authorities have tested them first. A British official confirmed the report to Bloomberg the next day. It isn't a law, and it doesn't stop the UK testing models forever — it changes the order in which things happen: America first, everyone else after.

That's the entire official explanation, from an unnamed senior administration official. In plain English: because the companies are American, Washington believes it should look first. Whether you find that reasonable probably depends on whether you think of AI safety testing as a shared international project, or as each country protecting its own patch first and sharing what's left over.

This isn't hypothetical. Anthropic released a new model, Claude Mythos 5.1, on 1 September 2026, and restricted it to a set of US organisations. On its own announcement, the company said:

AISI's director, Henry de Zoete, says no — the institute still has 'prerelease access to some of the world's most capable models' and tested OpenAI's newest model, GPT-6 Astra, before it came out. A UK government spokesperson backed that up, saying the institute 'continues to work closely with the US government, international partners and leading AI companies' and that safety risks 'do not stop at national borders'. Both statements are true as far as they go. What they don't fully answer is the narrower, more uncomfortable question: does the UK now see new frontier models on the same timetable as the US, or a step behind it? On the Mythos 5.1 evidence, it's a step behind — at least for that one model, at least for now.

The timing isn't a coincidence. This news broke the same week Prime Minister Andy Burnham was at the UN General Assembly in New York, pitching AISI as working 'hand in glove' with the US and the big AI labs, and calling for 'a single set of global principles and standards' on AI. President Trump, speaking at the same UN gathering, rejected that kind of coordinated approach outright, calling it a 'globalist scheme'. Two allied governments, in the same week, describing two quite different visions of how AI safety testing should work — and the model-access story is the clearest evidence yet of which vision is actually shaping what companies do.

There's a glossary worth keeping handy here, because the jargon is doing a lot of the confusion:

What to actually watch: on 13 October, MPs on the Commons Business and Trade Committee will question OpenAI, Anthropic, Google DeepMind and Meta directly on exactly this kind of access gap, with AISI's own director appearing the same day. That hearing, not this week's headlines, is where you'll find out whether Parliament thinks a voluntary, goodwill-based testing system can survive one ally deciding to go first.