China’s leading artificial intelligence companies have publicly disclosed model-specific safety test results for only 3.6% of their model releases, according to a new report by research firm SemiAnalysis, raising fresh questions about transparency in one of the world’s two AI superpowers.

The California-based firm reviewed 857 models released between 2021 and September 15 by nine leading Chinese developers: Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax and StepFun.

It found that just 31 releases, or 3.6%, had a published safety-evaluation result that could be matched to a specific model. Only nine, or 1.1%, had such results available at or before launch. For 813 releases, researchers found no safety disclosure at all, though companies may have run tests privately.

What counts as disclosure

SemiAnalysis defined disclosures narrowly: specific results tied to a named model, covering tests of harmful output, jailbreak resistance, toxicity, privacy, refusal behaviour or dangerous capabilities. General claims that a model had been safety-trained or evaluated did not count.

The report also noted that no major Chinese developer has released a frontier text model with publicly disclosed dangerous-capability tests spanning cyber, biological and loss-of-control risks. It did not provide comparable figures for American developers, though firms such as OpenAI, Anthropic and Google DeepMind have published safety reports, system cards or model cards for some major frontier-model launches.

Growing global concern

The findings land as security incidents involving autonomous AI agents have intensified worldwide debate over whether companies should slow down to build safer models. Australia said last month that an OpenAI agent breached a government health portal.

Reuters reported last week that Chinese AI agents had demonstrated an ability to deceive users, evade restrictions and conceal failures in tests, echoing concerns raised about advanced American systems. China has published an AI Safety Governance Framework identifying risks such as models acquiring system permissions without authorisation or deceiving evaluators, but SemiAnalysis says it imposes no mandatory duties linked to model capability, with binding rules focused on applications rather than frontier risk assessments.