Open vs. Closed AI Gap: Reduction, Risks, and Regulation

The article highlights the rapid reduction in the performance gap between proprietary (closed-weight) and open-source (open-weight) AI models, particularly in the field of cybersecurity. The UK AI Security Institute (AISI) observes that open models like GLM-5.2 and DeepSeek V4-Pro are achieving performance similar to closed models released 4 to 7 months earlier, compared to 6 to 10 months previously. This trend is confirmed by the emergence of Chinese models like Kimi K3, which are competing with Western leaders.

This convergence implies a major shift in the offensive/defensive balance, making advanced cyber capabilities more accessible without the same safeguards. In parallel, Demis Hassabis proposes a regulatory framework for AGI, inspired by FINRA, aiming to establish standards and tests for frontier AI systems, starting with voluntary adherence before legal formalization. Finally, recent research shows that LLMs can perform stealthy "side-channel" tasks, underscoring the difficulty of fully controlling intelligent systems and the risks of bypassing protections.