SemiAnalysis: of the 857 models published by Chinese AI giants, only 3.6 per cent were accompanied by security results

Alibaba, ByteDance, DeepSeek and Moonshot almost never publish security assessments, and none of them have disclosed any tests of dangerous capabilities for a state-of-the-art model.

California-based firm SemiAnalysis reviewed 857 models posted online between 2021 and September 15, 2026 by nine major Chinese players: Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.ai, MiniMax and StepFun. Only 31 of them, or 3.6%, had a security assessment result published by the developer. And only 9 (1.1%) had them at the time of their release or before; 813 versions, nearly 95% of the total, have never been accompanied by any such information.

A strict definition

The study only uses numerical results linked to a specific model: tests for harmful content, jailbreak resistance, toxicity, privacy, refusal or dangerous abilities. General statements about safety training were not counted. When results do exist, they often come after the fact: half of the 31 cases were documented after discharge, with a median time of 42 days and up to 349 days for DeepSeek-R1.

According to the report, no major Chinese developer has yet published dangerous capability tests (cyberattack, biology, loss of control) for a state-of-the-art text model, while OpenAI, Anthropic, and Google DeepMind document their major releases. China's governance framework does identify risks such as unauthorized permission acquisition or deceptive behavior, but does not impose obligations related to model capabilities.