The British AI Security Institute (AISI) has found that open-weight AI models are now closing the performance gap with top proprietary systems in cyber capabilities. According to the institute, models like GLM-5.2 and DeepSeek V4-Pro have reached a level that closed frontier models achieved four to seven months earlier. This marks a significant shift in the competitive landscape of AI cyber capabilities.
AISI tested these models using two methods: the 'Narrow Cyber Tasks' benchmark, which includes 70 tasks across four difficulty levels, and 'Cyber Ranges,' which simulate autonomous cyber capabilities in networks. In the Narrow Cyber Tasks, GLM-5.2 matched the performance of Opus 4.6 from February 2026, placing it about four months behind. DeepSeek V4-Pro performed at the level of Opus 4.5, released in November 2025. In the Cyber Ranges test, GLM-5.2 reached a level similar to Opus 4.5, while DeepSeek V4-Pro fell below Sonnet 4.5.
AISI notes that the gap in Cyber Ranges is wider, at around seven months, and treats the result as weaker evidence due to fewer test scenarios. The tests also cannot determine whether a model fails due to a lack of cyber capabilities or an inability to sustain planning across long, complex attacks. The institute also highlights that open models are significantly cheaper to run, with costs for a 100-million-token Cyber Range test dropping from $85 for Opus 4.5 to just $1.19 for DeepSeek V4-Pro.
Source: thedecoder
