News-July-18-2026
July 18, 2026 — The United Kingdom's AI Security Institute reported that recent open-weight AI models were narrowing the cyber-capability gap with frontier closed models, making open-weight model evaluation a live AI-governance issue.[1]
Contents
UK AI Security Institute says open-weight cyber capabilities are closing on frontier models
The AI Security Institute said it has tracked frontier-model cyber capabilities since 2023 and found that the most capable open-weight models it evaluated still lagged leading closed frontier systems.[1] The institute reported that recent open-weight models lagged the cyber capabilities of frontier closed models by roughly four to seven months, compared with a six-to-ten-month gap through most of 2025.[1] The institute said its open-weight-model evaluations were largely unimpeded by safeguards and that DeepSeek V4-Pro refusals on narrow cyber tasks were easy to circumvent in testing.[1] The findings matter for AI governance because open-weight releases can distribute frontier-adjacent cyber capability beyond a single hosted provider's access controls.[1]