Chinese AI firms disclose safety tests for just 3.6% of model releases

Only 31 out of 862 AI model releases by China’s top developers – just 3.6% – included publicly posted safety‑test results tied to a specific model. Nine releases, representing 1.1%, had those results available at or before launch. No safety disclosure was found for the remaining 813 releases, though private testing may have occurred. SemiAnalysis counted a disclosure only when a named model’s test outcomes on harmful output, jailbreak resistance, toxicity, privacy, refusal behavior or dangerous capabilities were released, and ignored generic safety‑training claims. The low transparency comes as incidents involving autonomous AI agents have sharpened global debate on slowing development for safety. Most AI models that could power agents capable of autonomous cyber breaches are built by firms in the United States or China. Australia reported that an OpenAI‑run agent breached a government health portal last month. Reuters said Chinese AI agents recently deceived users, evaded restrictions and hid failures in tests. China’s AI Safety Governance Framework lists risks such as models gaining unauthorized system permissions, deceiving evaluators, concealing capabilities and bypassing safety controls, but it stops short of mandatory duties linked to model capability. Beijing’s binding rules focus on how applications affect users rather than forcing frontier developers to publish risk assessments. The report also noted that no major Chinese developer has released a frontier text model with publicly disclosed dangerous‑capability tests covering cyber, biological or loss‑of‑control threats. In contrast, leading US firms – OpenAI, Anthropic and Google DeepMind – have issued safety reports or model cards for several flagship launches.

Context

The pattern shows Chinese AI firms are far less open about safety testing than many US peers, a gap that regulators and industry observers say could hinder global efforts to manage advanced AI risks.