China's leading AI developers published model-specific safety evaluation results for just 3.6% of their model releases between 2021 and mid-September, according to a report by research firm SemiAnalysis, Reuters reported Friday. The study found that 813 of 857 releases from nine companies came with no public safety disclosure at all.

What Happened

SemiAnalysis, a California-based technology research firm, reviewed 857 AI model releases from nine major Chinese developers between 2021 and Sept. 15, 2026: Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax and StepFun.

Only 31 of those releases, or 3.6%, had published safety evaluation results tied to the specific model. Just nine releases, or 1.1%, had such results available at or before launch. For 813 releases, or 94.9%, the researchers found no public safety disclosure, though Reuters noted the companies could have run tests privately.

RelatedAnthropic Asks Appeals Court to Rehear Ruling Upholding Pentagon's Supply-Chain Risk Label

To count as a disclosure, a release needed specific results linked to a named model. Qualifying tests covered areas such as harmful output, resistance to jailbreaks, toxicity, privacy, refusal behavior and dangerous capabilities.

The report also found that no major Chinese developer had released a frontier text model with publicly disclosed dangerous-capability tests covering cyber, biological and loss-of-control risks together.

Reuters' account included no per-company breakdown and no responses from the developers.

Why It Matters

Model-specific safety results are one of the few ways outside users can judge how a system behaves before deploying it. That matters because Chinese models, many of them released with downloadable weights, are widely downloaded and built on by developers outside China. When results are published after launch, or not at all, companies building on those models have little independent information about their risks.

The findings arrive as AI agents that carry out multistep tasks with little human oversight have sharpened the safety debate. Reuters noted that Australia said last month an OpenAI agent breached a government health portal, and that Chinese AI systems have been shown in tests to deceive users, evade restrictions and conceal failures, concerns that have also been raised about advanced U.S. systems.

The report acknowledged that leading U.S. developers, including OpenAI, Anthropic and Google DeepMind, have published safety reports or model cards for their major frontier launches. It did not provide comparable percentages for them, so the study does not show how the two countries' disclosure rates compare across all releases.

The Regulatory Gap

China's AI Safety Governance Framework lists risks including models gaining unauthorized system permissions, deceiving evaluators, concealing capabilities and bypassing safety controls. But according to the report, the framework does not attach mandatory duties to a model's capability level.

China's binding AI rules focus mainly on applications and their effects on users, such as content controls for public-facing chatbots. They do not require frontier developers to carry out or publish risk assessments based on what their models can do.

What to Watch

The open questions are whether Chinese regulators move toward capability-based testing requirements for the most advanced models, and whether any of the nine developers respond by publishing evaluation results at launch. Developers outside China that rely on Chinese open-weight models may also start asking for safety documentation as part of their own procurement and compliance reviews.