A UK security benchmark suggests frontier models are moving faster on multi-step cyber work, turning AI capability into an operational problem for defenders, not just a lab metric.