UK AI Security Institute finds Anthropic's Mythos model excels at multi-step cyberattacks