Lyptus Research: AI Offensive Cyber Capability Doubling Every 9.8 Months, Accelerating to 5.7 Months Post-2024

Lyptus Research·high signal

Lyptus Research applied the METR time-horizon methodology to offensive cybersecurity across 15 AI systems spanning 2019-2026, finding capability doubles every 9.8 months overall and every 5.7 months on a 2024+ fit — grounded in a new human expert study with 10 professional security practitioners. GPT-5.3 Codex and Opus 4.6 now achieve 50% success on offensive security subtasks requiring approximately 3 hours of professional expert time, while open-weight GLM-5 trails the closed-source frontier by only 5.7 months, suggesting offensive-cyber capability diffuses to open models on alarmingly short timelines.