UK AI Safety Institute Exposes Major Security Flaws in Leading AI Models