The United States has restored the United Kingdom's access to unreleased American AI models for pre-release safety testing, reversing a policy that had cut off cooperation between the two countries' AI safety bodies.
The freeze lasted 17 days, imposed in late September and lifted on 11 October, after the Trump administration asked US developers to withhold pre-release models from the UK until domestic evaluations were finished.
The UK's AI Security Institute screens advanced models before they reach the public, checking for risks such as assistance with cyberattacks or chemical and biological weapons development.
During the freeze, access to models from developers including OpenAI, Anthropic and Google DeepMind was effectively restricted, and at least one model release reportedly became to US organisations only, illustrating the practical bite of the policy.
UK AI minister Kanishka Narayan had pushed for the reversal.
The cooperation dates to November 2023, when the UK and US stood up parallel AI Safety Institutes, followed by an April 2024 memorandum committing both sides to joint model evaluations and research, including at least one joint test of a publicly accessible model; the two institutes had already run a joint pre-deployment evaluation of OpenAI's o1 model.
The reversal does not establish a binding regulatory regime, since developer cooperation still rests on voluntary access agreements rather than legal obligation.
