Open-weight models GLM-5.2 and DeepSeek V4-Pro match frontier cyber performance

Open-weight models GLM-5.2 and DeepSeek V4-Pro have achieved performance levels that match frontier cyber capabilities from just four months ago. This development marks a significant reduction in the gap of cyber capabilities, which previously ranged from six to ten months at the start of 2025. The British AI Security Institute reports that the effectiveness of safety measures on these open models has been largely ineffective. This follows a trend where open-weight models have rapidly closed performance gaps in a short timeframe, indicating a shift in the landscape for AI-driven cybersecurity solutions. Practitioners can now leverage these models for advanced cyber operations at a fraction of the previous cost, as highlighted by The Decoder.