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

Anthropic's recent release of the Mythos Preview model, now available to a select group of industry partners, marks a significant development in AI-driven cybersecurity tools. The UK government's AI Security Institute (AISI) conducted an evaluation of Mythos, highlighting its proficiency in chaining multiple cyber-attack tasks, which distinguishes it from other recent frontier models. While Mythos performs similarly to its peers on basic cybersecurity tasks, its ability to integrate complex, multi-step attack sequences allows it to tackle sophisticated infiltration scenarios, a task that would typically take a skilled human operator around 20 hours. This evaluation is critical for the AI/ML community as it provides independent verification of Anthropic's claims and illustrates the evolution of AI models in addressing cybersecurity challenges. AISI's Capture the Flag challenges have indicated a steady improvement in AI performance, with Mythos achieving a high success rate in Apprentice-level tasks. However, when compared to other emerging models like GPT-5.4 and Codex 5.3, the incremental advancements do not entirely justify the stringent access restrictions Anthropic has imposed. As the cybersecurity landscape evolves, the insights gained from tools like Mythos could shape future strategies for both AI development and defending against cyber threats.