Anthropic's unreleased Mythos LLM has demonstrated remarkable cybersecurity capabilities, completing a complex 32-step corporate network attack simulation in 3 out of 10 attempts. Third-party evaluation by AISI confirms the model's strength, noting no diminishing returns even at 100M token budgets ($12,500 per attempt). This creates a 'proof of work' dynamic in security: defenders must spend more tokens finding exploits than attackers spend exploiting them. Key implications include the continued importance of open source software (pooled token spending makes OSS more defensible), and the emergence of a three-phase agentic development cycle: development, code review, and hardening. Security costs are now effectively set by the market value of exploits.
Table of contents
Is security spending more tokens than your attacker?Sort: