UK gov’s Mythos AI tests help separate cybersecurity threat from hype (arstechnica.com)

Anthropic's recent release of the Mythos Preview model, now restricted to select industry partners, highlights a significant advancement in AI's capabilities for cybersecurity tasks. The UK government's AI Security Institute (AISI) independently evaluated Mythos, noting its competence in handling multi-step cyber-attack strategies that are crucial for infiltrating complex systems. While Mythos's performance in individual cybersecurity tasks aligns with other leading models, its standout ability lies in chaining together intricate sequences of attacks, demonstrating a practical application that extends beyond mere task completion. AISI's Capture the Flag (CTF) assessments reveal that Mythos achieved over 85% success in low-level CTF tasks, marking a notable high point in the competition among AI models. However, competitors like GPT-5.4 and Anthropic’s own Opus 4.6 show similar performance levels across various challenges. The unique advantage of Mythos surfaces in the advanced test dubbed “The Last Ones,” designed to mimic prolonged data extraction operations, which involve complex movement across diverse network segments. This capability not only signifies a leap in AI's operational potential but underscores the need for careful deployment in cybersecurity, balancing innovation with the risks of misuse.