The UK's AI Security Institute (AISI) reported Friday that Chinese open-weight AI models are closing the capability gap with US frontier models from OpenAI and Anthropic faster than expected. The agency said Chinese developers were six to 10 months behind in 2025, but that gap has fallen to as little as four months.

The findings follow Moonshot's release of Kimi K3, China's largest AI model, which benchmark results suggest outperforms Anthropic's Claude Opus 4.8 and OpenAI's GPT 5.5 on coding and agent tasks. Meanwhile, Beijing-based Z.ai's GLM-5.2 performed comparably to top US models on hacking tasks while using fewer tokens and at lower cost.

AISI warned that the narrowing gap between open and closed models could imperil global cyber systems, shortening time for organisations to secure defences. Once open-weight models are released, safety controls like access restrictions and model withdrawal are lost permanently. Researchers caution that widely available, highly capable AI could autonomously identify vulnerabilities in critical infrastructure.