UK AI tests Mythos cybersecurity capabilities

The UK's AI Security Institute (AISI) has evaluated Anthropic's Mythos AI model, finding it capable of chaining cybersecurity tasks to infiltrate systems. While Mythos performed comparably to other advanced models on individual tasks, it uniquely succeeded in completing a complex, multi-step infiltration challenge called "The Last Ones." This marks a significant advancement, demonstrating the AI's potential to autonomously attack weakly defended enterprise systems. However, the AISI cautions that real-world defenses and the model's limitations in more difficult scenarios still need consideration, emphasizing the need for AI-driven hardening of system protections.