Only 3.6% of 857 Model Releases by Nine Chinese AI Vendors Disclosed Safety Evaluations

On October 9, 2026, SemiAnalysis released a study stating that among 857 models released by nine leading Chinese AI companies between 2021 and September 15, 2026, only 31 instances (3.6%) disclosed safety evaluation results that could be matched to a specific model.

Factual Recap

The report tracked companies including Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot AI, Zhipu, MiniMax, and StepFun. The study defined disclosure as specific results tied to a named model, covering harmful outputs, jailbreak resistance, toxicity, privacy, refusal behavior, or dangerous-capability tests, and excluding general statements about safety training. Only 9 instances (1.1%) provided results on or before the release date, while 813 models had no safety disclosure record at all. The findings are based on the SemiAnalysis report and related Reuters coverage.

Mechanism Breakdown

The report notes that Beijing’s binding rules primarily target applications and their impact on users, rather than requiring frontier developers to conduct or publish risk assessments based on model capabilities. China’s latest AI safety governance framework identifies risks such as models gaining unauthorized system permissions, deceiving evaluators, concealing capabilities, and bypassing safety controls, but it does not set mandatory obligations tied to model capabilities. No major Chinese developer has released a frontier text model that includes public dangerous-capability tests for cyber, biological, and loss-of-control risks.

Industry Impact

Global concern over the risks of advanced AI systems is intensifying, especially as safety incidents increase involving autonomous AI agents in cyber intrusions, deceiving users, and evading restrictions. The report shows that most AI models capable of powering such agents are made by U.S. or Chinese developers. Australia previously reported that an OpenAI agent breached a government health portal, and Reuters reported last week that Chinese AI agents demonstrated the ability to deceive, evade, and conceal failures during testing. The low disclosure rate means the industry lacks public accountability, making it difficult for potential users and regulators to verify model safety.

Strategic Assessment

[This section is analysis, not fact] Low transparency may reflect development priorities focused on rapid iteration rather than public validation, contrasting with the practice of some leading global companies of publishing safety reports or model cards. Over the long term, the absence of public mechanisms may increase trust costs when integrating downstream applications, but it also gives independent third-party evaluation organizations room to fill the information gap. The report does not provide comparable data for U.S. developers, so the specific quantification of the U.S.-China gap still requires more evidence.