South China Morning Post reported on November 15 that major Chinese artificial intelligence model companies have begun restricting the network attack capabilities of their most advanced open-source weight models. Tencent stated this month that its latest Hy4 model has removed network attack functions while retaining defensive capabilities, though the open-source weight version has not yet been released. Similarly, Moonshot AI took comparable measures with its Kimi K3.1 Pro model in October.
These models are released with open weights, allowing anyone with appropriate hardware to download and modify them, making it difficult to prevent misuse through usage monitoring as is done with closed-source models. As a result, companies have shifted toward curating training data, eliminating attack-related capabilities at the source. Analysts note this approach remains an emerging practice rather than an industry standard.
Moonshot AI acknowledged the trade-offs involved: the filtered model scored 81% on internal cybersecurity benchmarks, compared to 90.5% for the unrestricted version. Analysts suggest that balancing openness, performance, and security may require compromises, though there is currently no evidence of broader competitive disadvantage.
Original: toutiao.com/article/1881890871569420/
Disclaimer: The views expressed in this article are those of the author alone.