National AI Security Institute Benchmarks Model Skills On Cyber Tasks

A newly published report from a national AI Security Institute presents detailed experiments measuring how modern AI systems handle cyber operations tasks at different skill levels, from basic apprentice exercises to challenges that previously required many years of experience. The findings show that success on apprentice-level cyber tasks has risen from under 9 percent in 2023 to around 50 percent in 2025, and for the first time a model was able to complete an expert-level exercise, giving policymakers and researchers a concrete data point on how quickly capabilities are advancing. The report stresses that safeguards are improving but still imperfect, noting that every evaluated system can be bypassed in some way, and calling for continued research into more reliable safety techniques so that organisations can benefit from AI assistance in areas like secure coding and threat analysis without increasing the risk of misuse.​

17-12-2025