“The UK AI Security Institute's mandate includes pre-release evaluation of frontier models for dangerous capabilities in biosecurity, cybersecurity, and loss of control.”