Michael Brunton-Spall boasts UK's AI Security Institute as the most advanced globally, detailing tests on Anthropic's Mythos model that autonomously attacks networks.