Question to the Cabinet Office:

To ask the Minister for the Cabinet Office, what access the AI Security Institute (a) requested and received to Claude Mythos 5.1 before its release to approved users and (b) has subsequently obtained; whether restrictions apply to that access; whether Anthropic gave any refusal, delay or restriction; what representations Ministers have made to Anthropic and the United States Government on this matter; and when the Government expects the Institute to complete an independent evaluation of that model.

The AI Security Institute (AISI) continues to collaborate closely with all industry partners, including Anthropic, and maintains long-standing arrangements to test models from major AI developers.

As I outlined in my speech on 12 October, the Government has been working closely with the US government and frontier labs to ensure the AI Security Institute continues to have access to the most powerful AI models before they are released.

AI scientists at the AI Security Institute continue to conduct rigorous evaluations of the latest generation of models to build a scientific understanding of the risks posed by the most advanced AI systems and develop technical mitigations against them, so that policy decisions are grounded in evidence.

Pre-release testing is a core part of the Institute's work to inform Government, but its remit extends well beyond this. Its researchers work on projects drawing on both pre-release and post-release testing across chemical and biological risks, human influence and loss of control. This research helps the UK and our partners make real-world improvements to AI security, for example by advancing the science of evaluations and AI monitorability.