Offensive Cybersecurity Time Horizons
Summary
Lyptus Research presents a robust analysis of offensive cybersecurity time horizons using METR's methodology, tracking AI capability growth across frontier models, token-budget scaling, and open-source diffusion. The report concludes that frontier capability has advanced significantly, with P50 around 3.2 hours for tasks requiring human-expert time, and highlights the adaptation buffer to open-weight models and implications for defense and policy.