Claude Mythos Cracks 73% of Expert Cyber Tasks No AI Could Solve Before

Anthropic's Claude Mythos Preview AI has become the first to successfully complete a full simulated corporate network attack, according to evaluations by the UK's AI Security Institute (AISI). The model achieved a 73% success rate on expert-level capture-the-flag tasks that were previously unsolvable by any AI. In a 32-step corporate network attack simulation, Mythos Preview completed an average of 22 steps, significantly outperforming Claude Opus 4.6. Anthropic also reported that Claude Mythos Preview can detect and exploit zero-day vulnerabilities when instructed. Due to its advanced capabilities, Anthropic is not releasing the model publicly and is instead using it for security research. The findings have prompted high-level discussions, including a meeting between US Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell with major bank CEOs regarding potential cyber risks. AISI recommends organizations prioritize foundational cybersecurity measures.

(Source：BeInCrypto)