The UK government said on 12 October 2026 that scientists at Britain's AI Security Institute are conducting rigorous evaluations of the latest generation of advanced AI models, after work with leading AI developers and the US government. The confirmation came in a Cabinet Office news release timed with AI Minister Kanishka Narayan's speech at Imperial College London's Base One incubator.
The same release sets out a wider package: putting frontier AI to work on government cyber defences, building on pilots that the government says have already identified and addressed more than 1,000 vulnerabilities; exploring stronger protections under the Cyber Security and Resilience Bill, including how serious AI incidents and near misses are identified and reported; and preparing with the National Cyber Security Centre for incidents involving autonomous AI agents.
Narayan said Britain should become one of the hardest countries in the world to attack using artificial intelligence, and urged model developers to listen when their scientists warn of serious dangers, investigate unexpected behaviour, and withhold systems they cannot adequately control. Cyber Minister Liz Lloyd linked the agenda to the Cyber Bill and to scaling AI-assisted protection of public services.
The AISI access point closes a story that began the day before. On 11 October, The Telegraph reported that the White House had reversed an earlier request that labs such as OpenAI and Anthropic withhold pre-release models from UK safety testers. Reuters relayed that report. On 12 October, multiple outlets quoted Narayan confirming that the UK would retain early access for safety testing. The GOV.UK release itself does not name OpenAI, Anthropic or a White House u-turn; it states that evaluations of the latest models are underway following work with developers and the US government.
No separate White House statement or lab confirmation was located during this run. The durable, first-party facts are the UK government's own announcement of active AISI evaluations and the associated cyber-resilience measures. Claims about a specific White House reversal remain attributed to The Telegraph and secondary reporting unless and until Washington confirms them.
For the AI ecosystem, the practical effect is that the UK's best-resourced national model-evaluation body says it is again testing the newest systems before public release, while London simultaneously hardens cyber and incident-response plans for agentic AI.
Taken together, the 12 October package is less a single access anecdote than a public signal that London wants AISI evaluations, cyber hardening and agent-incident planning to move in parallel. Readers should still separate the verified Cabinet Office statements from the earlier Telegraph account of White House deliberations until Washington speaks for itself.