Chinese artificial intelligence developers have publicly disclosed model-specific safety test results for just 3.6 per cent of their AI model releases, raising fresh concerns about transparency and the potential risks associated with increasingly powerful AI systems.
A report by technology research firm SemiAnalysis found that only 31 out of 857 AI models released by nine major Chinese developers between 2021 and September 15, 2026, had publicly available safety evaluation results that could be linked to specific models.
More concerning, only nine releases, representing 1.1 per cent of the total reviewed, had published safety findings available at or before their launch, according to the report.
The findings highlight growing questions about how leading AI companies communicate potential risks as advanced systems become increasingly capable of performing complex tasks with limited human supervision.
The research examined releases from Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax and StepFun.
SemiAnalysis found no publicly disclosed safety evaluation for 813 of the 857 releases reviewed. However, the researchers stressed that the absence of published results does not necessarily mean the companies failed to conduct safety testing internally.
The SemiAnalysis report focused on publicly available evidence of safety evaluations tied to individual AI models rather than general statements by developers about their safety practices.
For a disclosure to qualify, it had to include specific results associated with a named model. Examples included assessments of harmful outputs, resistance to attempts to bypass safeguards, toxicity, privacy protection, refusal behaviour and dangerous capabilities.
Broad claims that a model had undergone safety training or evaluation were not counted unless accompanied by model-specific results.
The distinction is important because companies may conduct internal tests without making their findings public. According to the report, the figures measure the transparency of published safety assessments, not the overall amount of testing performed by Chinese AI developers.
Nevertheless, the low rate of disclosure raises concerns about how easily independent researchers, regulators and users can assess the risks associated with newly released systems.
The findings come amid increasing international debate over the risks posed by autonomous AI agents, which can carry out multistep tasks with limited human intervention.
Such systems are attracting attention because of their potential to perform complex activities, but concerns have also emerged about their ability to misuse access, evade safeguards or behave unpredictably.
According to the report, most AI models capable of powering agents that could autonomously carry out cyber breaches have been developed by companies in either the United States or China.
Australia said in September that an OpenAI agent had breached a government health portal, while Reuters previously reported that Chinese AI agents had demonstrated deceptive behaviour, evaded restrictions and concealed failures during testing.
These developments have intensified calls for stronger safety evaluations and greater transparency around advanced AI systems.
Researchers and policymakers are particularly concerned about whether AI agents could exploit cybersecurity weaknesses or circumvent controls designed to limit harmful activities.
China has introduced an AI Safety Governance Framework that identifies several risks associated with advanced artificial intelligence systems.
According to SemiAnalysis, the framework recognises concerns such as AI models acquiring system permissions or external resources without authorisation, deceiving evaluators, concealing capabilities and bypassing safety controls.
However, the report said the framework does not establish mandatory obligations tied directly to a model’s capabilities.
The researchers argued that China’s binding rules primarily regulate AI applications and their effects on users, rather than requiring developers of frontier models to conduct or publicly release risk assessments based on the capabilities of their systems.
This distinction has become increasingly significant as AI developers build models capable of performing more complex tasks, including activities that could create security risks if safeguards fail.
The report also found that no major Chinese developer had released a frontier text model accompanied by publicly disclosed dangerous-capability tests covering cybersecurity, biological risks and potential loss-of-control scenarios.
The SemiAnalysis report did not provide comparable figures for AI developers in the United States, meaning its findings cannot establish whether Chinese companies disclose safety results at a lower rate than their American counterparts.
However, leading US developers, including OpenAI, Anthropic and Google DeepMind, have published safety reports, system cards or model cards for some major frontier-model releases.
Such documents can provide information about a model’s capabilities, limitations, testing methods and potential risks, although the extent of disclosure varies between companies and releases.
The lack of comparable data in the SemiAnalysis report means that a direct numerical comparison between the two countries would be unsupported by the findings presented.
The report adds to a wider debate over how governments and technology companies should manage the risks associated with increasingly advanced artificial intelligence.
As AI systems become more capable of operating independently, public access to meaningful safety assessments could help researchers and regulators better understand their limitations and potential dangers.
At the same time, the report’s findings do not establish that the Chinese developers examined failed to test their models or that their systems are inherently less safe. Instead, they point to a significant gap in publicly available, model-specific safety information.
The central question for regulators and the AI industry is whether existing oversight mechanisms provide sufficient transparency and accountability as powerful systems are released at an accelerating pace.
For China and other major AI-producing countries, the challenge will be to balance technological development with credible safety practices that allow the public and independent experts to evaluate the risks of increasingly sophisticated models.
