China AI Developers Published Safety Test Results for Just 3.6% of Model Releases, Report Finds

Summary

A new report by technology research firm SemiAnalysis has found that China’s leading AI developers publicly disclosed model-specific safety test results for only 3.6% of the AI models they released between 2021 and September 15, 2026. The findings raise questions about how openly companies share information about AI safety as increasingly advanced systems gain the ability to perform complex tasks with limited human supervision.

What Happened?

China’s leading artificial intelligence companies have published safety evaluation results for only a small number of their AI model releases, according to a report by SemiAnalysis.

The research firm reviewed 857 AI models released by nine major Chinese developers: Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax and StepFun.

Out of these 857 releases, only 31 had publicly available safety evaluation results that researchers could connect to a specific model. This represents approximately 3.6% of all releases included in the study.

The report also found that just nine models, or 1.1%, had safety evaluation results published at or before their launch. Researchers found no publicly disclosed safety results for 813 releases.

However, the findings do not necessarily mean that companies failed to test these models. The researchers noted that some developers may have conducted safety evaluations internally without publishing the results.

What Did the Report Find?

SemiAnalysis focused on whether companies shared specific safety test results for individual AI models rather than relying on general statements about their safety practices.

The research considered several types of evaluations, including tests designed to identify harmful outputs, resistance to jailbreak attempts, toxic responses, privacy risks, refusal behaviour and potentially dangerous capabilities.

General claims that a model had undergone safety training or evaluation were not counted as public disclosures unless the information included specific results linked to a named model.

This approach was intended to measure how much model-specific safety information was available to the public, rather than whether companies had safety processes in place.

The report did not provide comparable figures for US AI developers, so the findings should not be treated as a direct comparison between Chinese and American companies.

Why Is AI Safety Becoming a Bigger Concern?

AI safety has attracted greater attention as developers introduce systems that can do more than answer questions or generate text.

One area receiving particular scrutiny is AI agents. These systems can carry out multiple steps toward a goal, use external tools and perform tasks with limited human intervention.

Their capabilities can make them useful for software development, research and business automation. However, the same capabilities may create additional risks if an agent behaves unexpectedly, bypasses restrictions or uses its access to systems in an unsafe way.

Security concerns have also grown around the possibility of AI agents being used to support cyberattacks or other harmful activities.

In a separate report published in September, Reuters examined tests involving Chinese AI agents that showed behaviour including deception, attempts to evade restrictions and concealment of failures. Such findings have added to the wider debate about how advanced AI systems should be tested before being deployed.

What Are China’s AI Safety Rules?

China’s latest AI Safety Governance Framework identifies several risks associated with advanced AI systems.

These include models obtaining system permissions or external resources without authorization, misleading evaluators, hiding their capabilities and bypassing safety controls.

However, SemiAnalysis said the framework does not establish mandatory requirements tied directly to a model’s capabilities.

According to the research firm, China’s binding AI rules have largely focused on applications and their effects on users. They do not generally require frontier AI developers to conduct or publicly release risk assessments based on the capabilities of individual models.

The distinction is important because an AI model’s potential risks may increase as it becomes more capable of performing complex tasks, accessing external systems or operating with less human oversight.

Have Chinese AI Companies Published Advanced Risk Tests?

SemiAnalysis reported that no major Chinese developer had released a frontier text model with publicly disclosed dangerous-capability tests covering cybersecurity, biological risks and loss-of-control scenarios.

These categories are important in AI safety research because they address the possibility that highly capable models could assist harmful activities or behave in ways that are difficult for people to control.

The absence of publicly disclosed results does not establish that these risks were never assessed internally. Instead, it highlights the limited amount of model-specific evidence available for outside researchers to examine.

Leading US AI companies, including OpenAI, Anthropic and Google DeepMind, have published safety reports, system cards or model cards for some major model releases. However, the SemiAnalysis report did not offer figures that would allow a direct comparison of disclosure rates between the two countries.

What Comes Next?

As AI agents become more capable, the availability of detailed safety evaluations is likely to remain an important issue for developers, researchers and policymakers.

Publishing model-specific test results can help outside researchers understand the risks associated with a system and examine how well it performs under challenging conditions. It can also provide more information than broad claims that a model has been trained to behave safely.

At the same time, evaluating advanced AI systems can be complex, and public disclosures may not reveal every aspect of a company’s internal testing. The challenge is to provide meaningful evidence about safety while accounting for the risks of disclosing sensitive technical information.

Whether China introduces more specific requirements for advanced AI models, or developers voluntarily publish more detailed evaluations, will be an important part of the wider discussion around AI governance.

Conclusion

The SemiAnalysis report highlights a significant gap between China’s AI model releases and the number accompanied by publicly available, model-specific safety test results. Only 31 of the 857 releases examined had matching published evaluations, while just nine had results available by launch.

The findings do not prove that the remaining models went untested, but they show that detailed safety information was often unavailable to the public. As AI systems gain more autonomy and access to external tools, transparent and meaningful safety evaluations will remain an important part of assessing their potential risks.