AI security in China: of 857 models published by nine laboratories, only 3.6 per cent have published security results

SemiAnalysis has scrutinised announcements from ByteDance, Alibaba, DeepSeek, Moonshot and Zhipu: Beijing is taking a very cautious line on the risks, but is regulating usage far more strictly than the border controls.

Whilst Washington debates whether or not to slow down the AI race in the name of rivalry with China, SemiAnalysis set out to investigate what Chinese players are actually doing. Its analysts identified 857 model releases between 2021 and 15 September 2026 from nine developers: four giants (ByteDance, Alibaba, Tencent, Baidu) and five start-ups (DeepSeek, Moonshot, Zhipu, MiniMax, StepFun). Only 31 of these, or 3.6 per cent, were ever accompanied by a security assessment result published by the developer, and only 9 (1.1 per cent) at the time of release. For 16 others, the documentation arrived after the fact, with a median delay of 42 days.

A very cautious stance, with speed as the priority

On paper, however, Beijing’s stance is one of the most alarmist: the AI 3.0 security governance framework, published on 14 September by TC260, refers to recursive self-improvement and establishes a risk category for models that circumvent their safeguards, deceive their evaluators or refuse to be shut down. Yet this same document places innovation at the forefront of its principles; the major AI law was shelved in 2025, and the ‘AI+’ plan aims for 70 per cent penetration of AI agents by 2027.

According to SemiAnalysis, China primarily regulates what systems say and do to users (content labelling, rules on AI companions, on minors or on agents), without any obligations triggered by a model’s computing power or capabilities. The regime is therefore said to be stricter than any other when it comes to content, but more flexible than the European Union or California regarding state-of-the-art models.