Open Weight Models Lag Frontier Cyber Capabilities by 4-7 Months
Open-weight GLM-5.2 and DeepSeek V4-Pro matched Claude Opus 4.5.
Entities: AI Security Institute, DeepSeek V4-Pro, GLM-5.2
A new AISecurityInst post on X says its first public analysis of the gap between closed and open-weight AI models on frontier cyber tasks found a lag of 4 to 7 months with GLM-5.2 and DeepSeek V4-Pro, down from 6 to 10 months through most of 2025. In a widely shared quote-post on X, @scaling01 summarized that result by arguing open models are now about seven months behind frontier systems on UK AISI's long-horizon cyber ranges.
Combined views
19.3K
10 posts, first seen 1h ago
Commentary on X
@scaling01 Now do Kimi, even though this shows us nothing, both GLM-5.2 and prob also Kimi k3 has been potentially intentionally nerfed on cyber capabilities
@scaling01 open models 7mo from cyber parity. the floodgates just cracked.
@scaling01 These benchmark timelines are moving way too fast.
@AISecurityInst Great work, can't wait for K3!