China AI developers publish safety tests for just 3.6% of model releases, report finds

The SemiAnalysis report emphasizes a specific definition of safety disclosure: results tied to a named model, covering tests for harmful output, jailbreak resistance, toxicity, privacy, refusal behavior, or dangerous capabilities. This focused definition distinguishes concrete, verifiable testing from broader, less specific claims about a model being "safety-trained" or "evaluated," which were not counted. This distinction is critical for evaluating the transparency of safety efforts.

The findings emerge as global debate intensifies over security incidents involving autonomous AI agents, which can perform multi-step tasks with limited human intervention. The report notes that most AI models capable of powering such agents originate from either US or Chinese developers. The absence of public safety disclosures for the vast majority of Chinese models could contribute to broader concerns regarding the responsible development and deployment of advanced AI systems.

China's AI Safety Governance Framework identifies risks like models acquiring unauthorized system permissions or deceiving evaluators. However, SemiAnalysis points out that these rules primarily govern applications and their effects on users, rather than imposing mandatory duties on frontier developers to conduct or publish risk assessments based on a model's capabilities. This regulatory approach means that while risks are acknowledged, there is no requirement for developers to publicly demonstrate specific tests for dangerous capabilities, such as those related to cyber, biological, or loss-of-control risks, for frontier text models.