Question to the Cabinet Office:

To ask the Minister for the Cabinet Office, with reference to the Answer of 12 October 2026 to Question HL3587, how many of the vulnerabilities that the AI Security Institute identified pre-release in 2026 were fixed before release.

Through its relationships with the world’s leading AI developers, the AI Security Institute (AISI) is able to identify vulnerabilities in new models through pre-deployment testing. Testing results are then shared with developers so that they can improve their safeguards. For example, our biosecurity red-teaming across OpenAI’s ChatGPT Agent and GPT-5 revealed more than a dozen vulnerabilities, including universal jailbreak patterns. OpenAI addressed these through engineering fixes, policy updates and improved monitoring. By conducting safeguards red-teaming, we support developers to incrementally improve models through strengthening their overall defensive systems, rather than patching specific vulnerabilities.