China's top 9 AI developers published 857 models, only 4.7% published safety test results

"Reuters" reported that as the world becomes increasingly concerned about the risks that may arise from advanced artificial intelligence (AI) models, a research institution released a report pointing out that China's major AI developers only publish specific safety test results for a small number of developed AI models, accounting for only 4.7%.

SemiAnalysis, a technology research organization headquartered in California, examined 857 AI models published by China's nine largest AI companies from 2021 to September 15 this year. These companies include Alibaba, ByteDance, Tencent, Baidu, DeepSeek, MoonShot AI, Z. AI, MiniMax and StepFun.

The report pointed out that among the 857 models, only 31 have disclosed safety assessment results that can correspond to specific models, accounting for only 3.6%; another 9 models have published such results before or before publication, accounting for 1.1%. This means that there are 817 models that do not disclose safety-related information, but these companies may still test them internally.

SemiAnalysis said that its definition of "information disclosure" refers to specific results associated with a named model, including tests for harmful output, jailbreak resistance, toxicity, privacy, denial of behavior, or dangerous capabilities, and does not count general statements that a model has received security training or evaluation.

Most of the models that can support AI agents to conduct autonomous cyber attacks are developed by either American or Chinese developers.

Reuters reported last week that China's AI agents demonstrated in tests the ability to deceive users, circumvent restrictions and conceal failures, echoing concerns raised about advanced U.S. systems.