UK AI Security Institute finds open-weight models now trail closed frontier on cyber by months

The UK AI Security Institute said on July 17, 2026 that leading open-weight models now trail closed frontier models on cyber-capability tasks by four to seven months, down from a six-to-10-month gap a year earlier.

The finding, published by the UK government’s model-testing agency, quantifies how fast freely downloadable models are catching the best closed systems on offensive-cyber tasks — the capability regulators worry about most, because open weights cannot be recalled once released. AISI tested two open-weight leaders: GLM-5.2, released in June 2026, and DeepSeek V4-Pro.

On a 70-task narrow-cyber benchmark, GLM-5.2 matched Anthropic’s Claude Opus 4.6 (a four-month lag) and, on AISI’s ‘The Last Ones’ cyber range, reached as far as the older Claude Opus 4.5 (a seven-month lag). DeepSeek V4-Pro compared to Opus 4.5 at a five-month lag while costing about $0.28 per task against $12.50 for that closed model; GLM-5.2 ran at $6.12 per task versus $15.17 for Opus 4.6, according to AISI’s analysis.

The results measure narrow cyber benchmarks, not real-world attacks. A closing capability gap does not mean open-weight models can run end-to-end intrusions unaided; benchmark performance and operational offensive capability are not the same thing.

Still, the direction is what matters for policy. If the gap keeps narrowing, the window in which closed-model safeguards can meaningfully gate dangerous capabilities shrinks with it.