Only 31 of 857 Chinese AI models have security test results

China’s AI ecosystem is challenging the West in how fast it produces models, but the same speed is not visible in security documentation. According to SemiAnalysis’s October 2026 report, only 31 of the 857 Chinese models introduced from 2021 through September 15, 2026 have publicly available security test results. That is 3.6%; the story was first reported by Reuters on October 9, 2026.

Testing and publishing results are not the same thing

The report counts not whether models were tested, but whether the results were disclosed. This distinction is critical: the engineering blogs of names such as DeepSeek, Alibaba’s Qwen team, and ByteDance are known for red team and evaluation work. By contrast, the model card and system card tradition of leading U.S. labs has almost no equivalent in China. The report notes that nine institutions have published no public results, and that the silence around the remaining hundreds of models cannot be reduced to a single reason: some may not have tested, while others may have tested but not shared. So the picture measures accountability, not technical quality.

Why is the West looking right now?

The timing is no coincidence. Over the past two months, perceptions of Chinese AI have been pulled to two extremes. On one side is Nvidia’s share of the Chinese chip market falling to 34% and the rise of Huawei Ascend. On the other is Anthropic’s distillation defense against Zhipu and Moonshot, and its mid-October announcement that reasoning-token theft targeting Claude had fallen by 90-99%. In other words, while the capability of Chinese models is debated, the ‘nobody knows what’s going on inside’ file is also opened at the same table. Regulators, enterprise risk teams, and researchers are all looking at the same gap.

The economics of transparency

In the short term, a lack of transparency does not look like a cost. In a landscape where DeepSeek closed a $12 billion funding round in September 2026 at a $75 billion valuation and Qwen is gaining share in the token market with open weights, a closed box raises fewer questions. But in the long term, the bottleneck shows up in enterprise adoption: banks, hospitals, and the public sector turn to suppliers from whom they can request security documentation. The sovereign cloud debate in Europe rests on exactly this point.

A practical read for Turkey

For teams running models such as Qwen, DeepSeek, and GLM on their own servers, the absence of security documentation often makes the extra cost look like zero. The real cost appears the moment we do not know what the model does not know.

Source: WebTuring