What is UK AISI?

UK AISI (AI Safety Institute) is a UK government body focused on evaluating frontier AI models for safety risks. As of March 2026, Geoffrey Irving, a researcher at UK AISI, stated that the institute has consistently found vulnerabilities in models: ‘every time we did we jailbroke a model.’ The host also noted: ‘We’ve never failed to jailbreak a model. None of them are robust.’ Irving observed that ‘the strong jailbreaks are concentrated in very particular domains,’ with bio being a key example. Irving also stated that UK AISI is ‘shifting a lot of that work… to longer research collaborations that might… be post deployment or… go back further before deployment is finalized,’ in response to the increasing pace of model releases.