Offensive Cybersecurity Time Horizons

Lyptus Research applies METR's time-horizon methodology to offensive cybersecurity across 7 benchmarks, finding AI capability is doubling every 9.8 months overall — and 5.7 months since 2024. The most capable frontier models (GPT-5.3 Codex and Opus 4.6) now match human expert performance on 3+ hour offensive security tasks, and the researchers believe even these estimates understate recent progress due to fixed evaluation budgets.