AI Cyberattackers Advance: UK Warns of Rapid Threat Growth

AI cyberattackers are rapidly improving, with the difficulty of cybersecurity defenses they can crack doubling every four months. This interval is also shrinking. New benchmarks from the UK government's AI Security Institute, or AISI, show dramatic improvements in AI models' ability to perform end-to-end penetration tests. In November 2025, the difficulty of cyber tasks AI models could complete was doubling every eight months. By February of this year, this accelerated to 4.7 months. Newer models like Claude Mythos Preview and GPT-5.5 show even greater capability. This measure focuses on autonomous capability, not speed, by comparing AI success rates to human expert performance. This means AI models can sustain performance over multiple steps, maintaining context and recovering from failures. The UK government is concerned about these findings. A UK AI Minister stated these advances are translating into real risks for organizations with weak cyber defenses. This matters because the rapid improvement of AI in cybersecurity presents both significant threats and potential tools for defense.