Rapid acceleration of AI cyber attack capabilities “Defense difficulty doubles every 4 months”

The ability of AI models to perform end-to-end, multi-stage penetration testing comparable to that of human experts has improved significantly in recent months, according to new benchmarks published by the UK government's AI Security Institute (AISI). According to AISI, a research institute under the Department of Science, Innovation and Technology (DSIT), as of November 2025, the difficulty of cyber tasks that top AI models can perform has doubled approximately every eight months. However, in February of this year, the speed of performance improvement became even faster, reaching a level where the difficulty of tasks that AI models can solve doubles approximately every 4.7 months.