TL;DR
Washington’s ONCD (Office of the National Cyber Director) wants OpenAI and Anthropic to keep their newest models away from the UK’s AI Security Institute (AISI) until American government testers have been through them first. Anthropic seems to have gone along with it: AISI has not received Claude Mythos 5.1. The move threatens the early, pre-release access that underpins much of the UK’s claim to a lead on AI safety.
What Washington asked for
According to City AM, the request gives the US government first sight of any new frontier model, with the UK’s safety institute waiting its turn. A senior administration official framed it as a security step, saying Washington wanted American systems secured before they reached allies. “Because they’re American companies and this has been our policy with every new frontier model that comes out,” the official said.
Anthropic did not share Mythos 5.1 with AISI, explaining that access was initially restricted to US organisations. The company said it was “coordinating with the US government to expand access to a broader set of domestic and international partners as quickly as possible”.
AISI director Henry de Zoete conceded earlier this month that the institute had not had Anthropic’s latest model. He said AISI kept “strong relationships with all frontier AI developers”, and pointed out that it had tested OpenAI’s GPT-6 Astra ahead of launch.
Why it matters for the UK
Pre-release testing is the institute’s core offer. Its researchers probe frontier systems for dangerous abilities, such as helping with cyber attacks or biological threats, before the public gets them. If American testers must go first, the window AISI gets before launch can only shrink.
The timing is poor for AISI. Its chief technology officer, Jade Leung, is stepping back to a vice-chair role at the end of the month, and some staff have reportedly taken stress leave amid punishing testing schedules. At the UN this week, lab chiefs Sam Altman and Dario Amodei asked for shared safety standards, which the US administration turned down. Andy Burnham, meanwhile, has pitched the UK as a leader on global AI standards and wants them at the centre of the UK’s coming G20 presidency.
Looking forward
The UK’s safety model depends on access that American labs such as OpenAI and Anthropic choose to grant. If US agencies now decide when that access starts, the institute’s influence rests on Washington’s goodwill as much as on the labs’. The test will be the next major release: whether AISI sees it before launch, or only once American reviewers are done with it.