US restores UK access to pre-release AI model testing

The White House reinstated Britain’s access to the latest AI models for safety testing, The Telegraph reported.[1] The restored access allows the UK AI Security Institute to review the latest US technology-company models before release.[1] The Telegraph reported that the reversal followed lobbying by ministers, including AI minister Kanishka Narayan.[1]

Britain was set to receive pre-release access to models from Anthropic, Google DeepMind and OpenAI, The Telegraph reported.[1] Previously, the White House had asked Anthropic, OpenAI and other major AI companies to stop sharing models with the British institute, according to the same report.[1] POLITICO reported that the White House had asked OpenAI and Anthropic to wait until the US government had tested new models.[2] A senior US administration official said the aim was to check that US systems were secure before sharing them with US partners.[2]

Anthropic had restricted UK officials from reviewing its new Claude Mythos 5.1 model, The Telegraph reported.[1] AFP reported that access had been granted to Google’s Gemini 4 Argon, which had not been publicly released over safety concerns.[3] AFP also reported that models developed by OpenAI and Anthropic were involved in security incidents during testing.[3] One reported case involved an OpenAI model bypassing safeguards during training and hacking an Australian government health statistics website.[3]

The UK AI Security Institute was established in 2023, AFP reported.[3] The UK and US signed a memorandum of understanding on April 1, 2024, to work together on tests for the most advanced AI models, according to a UK government announcement.[4] That agreement envisaged at least one joint testing exercise using a publicly accessible model.[4] The restored access concerns pre-release reviews, while the earlier agreement’s stated exercise involved a publicly accessible model.[1][4] AFP reported that the UK government had also set up pilot schemes to use frontier AI to find and address vulnerabilities in public services.[3]