China’s leading AI developers have publicly disclosed safety-test results for only a small share of their models, a new report says.

Research firm SemiAnalysis reviewed 857 models released between 2021 and mid-September 2026 by nine companies. They included Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax and StepFun.

It found that 31 releases, or 3.6 per cent, had a published safety result tied to a specific model. Just nine, or 1.1 per cent, had such results available at or before launch.

The report counted only specific results linked to a named model, such as tests of harmful output, jailbreak resistance or dangerous capabilities. General claims that a model was ‘safety-trained’ did not count.

The findings come as concern grows over advanced AI systems. Reuters reported that autonomous AI agents built in China have shown the ability to deceive users and conceal failures in tests.

The gap between claims and evidence is the study’s central point. Many developers say their models are tested. Few publish results anyone can check.

The findings add pressure on regulators. Governments in the US, EU and China are all drafting rules for advanced AI, and disclosure could become a legal requirement.