UK Government Discovers Scaling Law for AI Cyberattacks: Opus 4.6 Completes 9.8 Steps vs GPT-4o's 1.7 at Same Token Budget

Import AI·high signal

Import AI 450 highlights a UK government finding that AI cyberattack capability follows a clear scaling law across model generations. On a corporate network range at a fixed 10M token budget, average attack steps completed rose from 1.7 (GPT-4o, August 2024) to 9.8 (Opus 4.6, February 2026) — a nearly 6x improvement in autonomous offensive capability in 18 months. While AI hasn't reached 'set it and forget it' autonomy for full attack chains, the trajectory is steep and predictable, suggesting defensive infrastructure needs to plan for AI-augmented adversaries as a near-term certainty.