UK government finds open weight models are closing the cyber capability gap with closed frontier models

8:00 AM · July 18, 2026

The UK's AI Security Institute published its first public comparison of open weight and closed AI models' cyber capabilities, finding that recent open models GLM 5.2 and DeepSeek V4 Pro perform similarly to frontier closed models released four to seven months earlier, a narrower gap than the six to ten months the institute measured internally through most of 2025. On the institute's narrow cyber task suite, GLM 5.2 matched the performance of Anthropic's Opus 4.6 and OpenAI's GPT 5.3 Codex, both released four months prior, while DeepSeek's V4 Pro matched Opus 4.5, released five months earlier. On longer autonomous cyberattack simulations the gap was somewhat wider. The institute noted the open models were also far cheaper to run and that their safeguards were easy to bypass, since no single party controls access once weights are public. It frames the lag as a preparation window, time during which defenders with access to the strongest closed models can act before comparable capabilities become freely available without matching safeguards. That window is narrowing just as frontier cyber capabilities themselves have jumped sharply, with Anthropic's Claude Mythos Preview and OpenAI's GPT 5.5 both posting the largest capability gains the institute has recorded since it began testing in 2023.