The Trump administration is asking American AI labs to hold back their newest models from British safety testers until the US government has reviewed them first. That’s a significant friction point for a UK institute that has built its entire value proposition around getting early access to frontier models.
According to Live Mint, the request came from the Office of the National Cyber Director and applies specifically to OpenAI and Anthropic, the two most closely watched US AI labs right now. A senior administration official framed it simply: these are American companies, and this has been policy for every new frontier model. Anthropic has already complied. It declined to give the UK’s AI Safety Institute access to Claude Mythos 5.1, saying the model was available only to a select group of US organizations.
What this means for the UK AI Safety Institute
The UK’s AISI, created in 2023, has made pre-release model testing its core function. It works on a voluntary basis with major developers, and that access has been central to Britain’s pitch as a serious player in international AI governance. So losing access to Anthropic’s latest model is not a minor administrative hiccup. It goes to the heart of what the institute is supposed to do.
AISI Director Henry de Zoete acknowledged the gap in a letter to a UK parliamentary committee. Still, he noted the institute retained access to some capable models and had tested OpenAI’s GPT-6 Astra ahead of its public release. That’s not nothing. But the Anthropic situation signals that the arrangement is no longer unconditional.
Why Washington is drawing a harder line
The policy shift reflects real concerns, not just political posturing. Recent testing has shown that some advanced models can carry out sophisticated cyber activities, including breaking into external systems. An Australian government website was reportedly among the targets during controlled testing. If frontier models carry that kind of capability, the US argument for reviewing them before sharing with any foreign partner, even a close ally like the UK, has some logic behind it.
The administration is also still working out its broader posture on powerful AI systems. Requiring a domestic review before international access is one way to buy time and maintain some visibility over what these models can do.
The bigger picture for AI safety cooperation
This creates a real tension. The UK has spent two years building a model of international AI safety collaboration, running joint evaluations and promoting shared standards. A UK government spokesperson said AI risks don’t stop at borders and no country can handle them alone. That’s a reasonable position. But it’s now running into a US administration that is thinking about frontier AI through a national security lens first.
For developers and teams tracking AI governance, the key questions now are:
- How long does US government review take, and does it create meaningful delays for international partners?
- Will this policy extend beyond the UK to other allied testing bodies?
- Does it affect how Anthropic and OpenAI structure their international go-to-market strategies?
So far this looks like a policy in formation rather than a fully settled framework. But the direction is clear: the US wants to be first in line, even ahead of allies.
