AI Is Getting Better at Security and It's Doing It Faster Than Expected
IT Pro, Thursday, May 14th, 2026
AI models are doubling their cybersecurity task capabilities every 4.7 months, faster than anticipated.
According to the UK's AI Security Institute, AI models are rapidly improving at handling complex security tasks, with capabilities doubling every 4.7 months since late 2024 significantly faster than the previously estimated 8-month doubling rate. The acceleration has been further boosted by new security-focused models like Anthropic's Claude Mythos and OpenAI's GPT-5.5, which have exceeded both previous doubling-rate trends.
The institute measures progress through a time-horizon benchmark that tests AI success rates on tasks of varying complexity, though it acknowledges the benchmark has limitations. Experts warn that while AI presents both opportunities and risks for cybersecurity, organizations must immediately invest in strong security baselines to build resilience as attackers increasingly leverage these advancing AI tools.