UK AI Security Institute: GPT-5.5 Achieves 71.4% on Expert Cyber Tasks, Completes 32-Step Network Attack End-to-End

Simon Willison / UK AISI·high signal

The UK AISI evaluation found GPT-5.5 scored 71.4% on expert-level cybersecurity tasks and completed a 32-step network attack range end-to-end — only the second model after Claude Mythos to do so. Tasks were developed with Crystal Peak Security and Irregular. Simon Willison highlighted this as evidence of a cybersecurity capability overhang across frontier models.