A report by technology research firm SemiAnalysis has raised questions about the public disclosure of safety testing by leading Chinese AI developers. Published findings covered 857 models released between 2021 and September 15, 2026, by nine companies: Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax and StepFun. The research was reported by Reuters and published by The Indian Express on October 11.
SemiAnalysis found that 31 releases, or 3.6% of the total, had published safety-evaluation results that could be matched to a specific model. Only nine releases, equivalent to 1.1%, had such results available at or before launch. The researchers found no public safety disclosure for 813 releases, although they noted that companies might have conducted tests privately. Their definition of disclosure included model-specific findings about harmful outputs, resistance to jailbreak attempts, toxicity, privacy, refusal behaviour or dangerous capabilities.
The distinction between testing and publishing results is important. A company may conduct internal evaluations without making the findings available for independent scrutiny. Public results can help researchers, users and policymakers understand what a model has been tested for and where its risks may remain. The report therefore highlights a transparency issue, but it does not establish that every model without published results was released without safety testing.
SemiAnalysis also said it had not identified a major Chinese developer that had released a frontier text model with publicly disclosed dangerous-capability tests spanning cyber, biological and loss-of-control risks. The report did not provide comparable figures for US developers. It described China's existing framework as addressing risks such as unauthorised system access, deception and bypassing safety controls, while arguing that binding requirements did not mandate capability-based risk assessments for frontier developers.