Question to the Cabinet Office:

To ask the Minister for the Cabinet Office, whether he has made an assessment of the adequacy of existing testing and oversight arrangements for advanced artificial intelligence models in light of recent increases in their capabilities; and whether he plans to introduce further safeguards.

In light of the capability developments demonstrated by the instances of unauthorised behaviour by AI models, AISI has reviewed its testing practices and made updates to increase the security of testing evaluations to reflect model capabilities. AISI recently completed pre-deployment testing of OpenAI’s newest model, GPT-6 Astra. We maintain daily contact with leading AI developers on AI security testing. Further detail is provided in Written Statement UIN HCWS314 'Artificial intelligence update'.