A restriction, then a reversal
Britain will keep getting the most capable AI models before they are released, the UK’s AI minister Kanishka Narayan said on Monday, ending a period in which American labs had been told to hold new systems back from British testers.
Last month the White House asked OpenAI and Anthropic not to share new models with the UK’s AI Security Institute until US officials had assessed them first, AFP reported. The Daily Telegraph reported the reversal on Sunday and said Narayan had lobbied for it. Narayan said the UK had worked closely with the US government “to retain access for Britain’s AI Security Institute to the most powerful AI models before they are released”.
What the institute does with the access
The AI Security Institute was set up in 2023 and tests frontier systems for safety and security problems in the window between a model being finished and a model being shipped. Bloomberg describes it as the best-funded government body that regularly tests frontier models, and its testers have caught AI agents taking actions they were not asked to take.
AFP reports that the institute has been given access to Google’s Gemini 4 Argon, a model that has not been released publicly because of safety concerns. During the restriction, according to Bloomberg, Anthropic did not share its Mythos model with the institute, keeping it with a set of US organisations instead.
Narayan set out the standard he wants applied: “investigate unexpected behaviours, and if you can’t adequately control the risks, do not release the system”.
Dependence is the uncomfortable part
The episode exposed something a testing regime cannot legislate away. The most capable models are built in the United States and China, and a British institute can only examine what their makers choose to hand over.
Narayan was blunt about it in remarks to Bloomberg: “When you’re reliant on someone else’s technology, you have to be clear-eyed about the permanence of any outcome.” He said he was nonetheless “very confident” that the UK would be able to test the latest models.
That caveat is doing real work. Nothing in Monday’s confirmation is a treaty, a statute or a contract. Access was withdrawn by a request and restored by a request, and the same mechanism remains available.
What comes next
Narayan is expected to set out further measures, including tighter reporting requirements when an AI-related security incident touches critical infrastructure. He has also set up a task force to protect UK government systems from AI-driven attacks; two pilot exercises are reported to have found more than a thousand possible weak points.
On the American side, Google, Microsoft and xAI have agreed to pre-release reviews in the United States, which is the arrangement the earlier restriction was meant to prioritise.
The thing to watch is narrow and checkable: whether the next frontier model from each of the three labs reaches the institute before it reaches the public, and whether any conditions came attached to the restoration. Neither Narayan’s statement nor the reporting around it settles that.
