Our evaluation of Claude Mythos Preview’s cyber capabilities

The AI Security Institute evaluated Claude Mythos Preview's cybersecurity capabilities, finding it capable of executing multi-stage network attacks and discovering vulnerabilities autonomously in controlled simulations. It became the first model to complete a full 32-step network attack scenario (3 of 10 attempts) and achieved 73% success on expert-level capture-the-flag challenges. Evaluations noted significant limitations: test environments lacked active defenders and real-world defensive tooling, so effectiveness against hardened systems remains unproven.