Research by SemiAnalysis indicates that leading Chinese AI developers rarely disclose safety testing results for their models. The study analyzed 857 models released between 2021 and September 15, 2026, by nine major companies, including Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax, and StepFun.
Out of the total models reviewed, only 31 cases, or 6.3 percent, featured published safety evaluation results linked to a specific model. Furthermore, only nine releases, representing 1.1 percent of the total, provided such data at or before the time of launch. No information regarding safety testing was found for 813 of the models.
The disclosure criteria used by researchers included tests for malicious content generation, jailbreak resistance, toxicity, privacy compliance, and refusal behavior. General corporate statements regarding safety training were excluded from the assessment. No major Chinese developer has publicly released results for advanced LLMs regarding dangerous capabilities.
Current Chinese AI safety regulations focus primarily on applications and their impact on users rather than imposing requirements on developers to conduct or publish risk assessments based on model capabilities. The existing management system does not mandate that developers disclose findings related to potential risks like unauthorized system access or deceptive behavior.
