AI cyber capability is speeding past earlier projections

**Executive summary:** The UK AI Security Institute reports that frontier AI models such as Claude Mythos Preview and GPT-5.5 have rapidly improved autonomous cyber and software capabilities, doubling performance on time-horizon benchmarks every few months, achieving near-100% success on certain limited benchmark tasks, and completing complex multi-step simulated cyber-range attacks; the institute warns current evaluation frameworks may soon be insufficient and cautions about uncertainty in future rates of progress and defensive effectiveness.