AI Security Institute
(asked on 22nd October 2026) - View Source
Question to the Cabinet Office:
To ask the Minister for the Cabinet Office, on how many occasions since 1 January 2026 the AI Security Institute has recommended changes to a model before its release that the developer did not make.
Answered by
Kanishka Narayan
Minister of State (Cabinet Office) (Jointly with the Department for Business, Innovation, Science and Technology)
This question was answered on 2nd November 2026
The AI Security Institute (AISI) is able to identify vulnerabilities in frontier models through pre-deployment testing. AISI does not recommend specific changes to models. Instead, testing results are shared with developers so that they can improve their safeguards. By conducting safeguards red-teaming, AISI supports developers to incrementally improve models by strengthening their overall defensive systems.