UK AISI: AI Cyber Capabilities Doubling Every 4.5 Months
The UK AI Safety Institute's independent assessment finds the length of autonomous cyber tasks frontier models can complete is doubling every 4.5 months, with the rate accelerating. Both Anthropic's Mythos and OpenAI's GPT-5.5 show major capability gains and appear constrained by token budget, not underlying ability—meaning the current ceiling is resource allocation, not fundamental capacity. Findings align with independent METR assessments released simultaneously.
Why It Matters
A 4.5-month doubling cycle means today's evaluation frameworks will be severely understating capability within two years. The upper bound on both models remains undetermined.