White House Restores UK AI Safety Institute's Access to Frontier Models
The White House has reversed its September decision to cut the UK AI Security Institute off from early access to frontier models built by Anthropic, Google DeepMind, and OpenAI, following diplomatic pressure from UK AI Minister Kanishka Narayan.
The White House has reversed its earlier decision to block the UK's AI Security Institute from accessing frontier artificial intelligence models developed by Anthropic, Google DeepMind, and OpenAI, restoring a channel that had been severed in September. The reversal follows sustained diplomatic pressure from the UK's AI Minister, Kanishka Narayan, according to reports.
The move restores the British agency's ability to examine advanced AI systems before they are publicly released, a practice known as early access or pre-deployment testing. Such access allows safety researchers to probe models for dangerous capabilities, including cyber-offensive potential, biosecurity risks, and the capacity for deception or autonomous action, before those systems reach the public.
The September decision had instructed the three leading AI companies to stop sharing their models with the UK body, which was established to evaluate the safety of frontier AI systems. The interruption raised concerns among researchers and policymakers about the ability of governments to independently assess rapidly advancing technology. It also threatened to complicate cooperation between the United States and the United Kingdom on AI governance at a time when both countries have emphasised the need for international coordination on safety standards.
Narayan's intervention appears to have been decisive. The UK minister, who holds the AI brief in the British government, engaged in diplomatic efforts that ultimately persuaded Washington to restore the arrangement. The reversal signals that the UK's role in AI safety testing remains valued by the US administration, despite earlier signals that access would be curtailed.
The restoration of access is significant for the global AI safety landscape. The UK AI Security Institute, formerly known as the AI Safety Institute, was among the first government bodies dedicated to evaluating frontier models. Its work has been watched closely by other nations considering similar oversight mechanisms. Losing access to models from Anthropic, Google DeepMind, and OpenAI would have left the institute unable to conduct independent testing on some of the most capable systems in development.
For the companies involved, the reversal means they can resume sharing their models with the UK body under the previous arrangement. Anthropic, Google DeepMind, and OpenAI have all previously cooperated with the institute, providing early access as part of voluntary commitments to safety and transparency. The interruption had forced them to choose between complying with the US directive and maintaining their relationship with the British regulator.
The episode highlights the delicate balance between national security concerns and international cooperation on AI safety. The Trump administration's initial move to restrict sharing may have reflected worries about technology leakage or a desire to keep frontier AI expertise within US borders. But the swift reversal suggests that such concerns were weighed against the diplomatic and strategic value of maintaining close ties with a key ally on AI governance.
It remains unclear whether the restoration of access comes with new conditions or limitations. Neither the White House nor the UK government has publicly detailed the terms of the renewed arrangement. The companies have not issued statements on the matter.
The decision also comes amid a broader debate over how best to regulate AI. While some argue for strict controls and government-led testing, others favour industry self-regulation and lighter-touch oversight. The UK institute's model, which combines government oversight with voluntary industry cooperation, has been cited as a potential template for other countries.
For now, the immediate effect is that the UK AI Security Institute can once again examine frontier models from the three leading developers. That access is likely to be crucial as AI systems grow more capable and the stakes of safety testing continue to rise.
8
