報告：中國AI開發商僅為3.6%的模型發布公開安全測試結果

愛德華多・巴普蒂斯塔報導By Eduardo Baptista

北京，10月9日（路透社）——研究公司SemiAnalysis在一份報告中表示，隨著全球對先進AI系統所帶來風險的擔憂升高，中國主要AI開發商僅對其一小部分發布模型，公開揭露特定模型的安全測試結果。BEIJING, Oct 9 (Reuters) - China's leading AI developers have publicly disclosed model-specific safety-test results for only a small fraction of their releases, a report by research firm SemiAnalysis said, as concerns about the risks posed by advanced AI systems increase worldwide.

總部位於加州的科技研究公司SemiAnalysis，檢視了中國9家主要AI公司在2021年至9月15日期間發布的857個模型；這些公司包括阿里巴巴、字節跳動、騰訊、百度、DeepSeek、Moonshot、Z.AI、MiniMax和StepFun。California-based SemiAnalysis, a technology research firm, reviewed 857 models released between 2021 and September 15 by nine leading Chinese AI companies — Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax and StepFun.

報告指出，其中31個發布模型、也就是3.6%，有公布可對應至特定模型的安全評估結果；僅9個、也就是1.1%，在發布時或發布前已有這類結果可供查閱。研究人員發現，813個發布模型沒有安全揭露資訊，不過相關公司可能已私下進行測試。It found that 31 releases, or 3.6%, had a published safety-evaluation result that could be matched to a specific model. Just nine, or 1.1%, had such results available at or before launch, it said. Researchers found no safety disclosure for 813 releases, though companies could have conducted tests privately.

SemiAnalysis表示，該公司將「揭露」定義為與具名模型相關的具體結果，包括有害輸出、越獄抵抗能力、毒性、隱私、拒絕行為或危險能力等測試；若只是泛稱某模型已接受安全訓練或評估，則不列入計算。SemiAnalysis said it defined disclosures as specific results tied to a named model — including tests of harmful output, jailbreak resistance, toxicity, privacy, refusal behaviour or dangerous capabilities — and did not count general claims that a model had been safety-trained or evaluated.

這些發現出爐之際，涉及自主AI代理的安全事件已使全球辯論升溫，焦點在於企業是否應該放慢腳步，以打造更安全的模型。自主AI代理是指能在有限人類介入下執行多步驟任務的系統。The findings come as security incidents involving autonomous AI agents — systems that undertake multistep tasks with limited human intervention — have intensified global debate over whether companies should slow down to build safer models.

能支援AI代理自主發動網路入侵的AI模型，絕大多數由美國或中國開發商製作。The vast majority of AI models capable of powering agents that could autonomously carry out cyber breaches are made by either US or Chinese developers.

澳洲上月表示，一個OpenAI代理入侵了政府健康入口網站。Australia said last month an OpenAI agent breached a government health portal.

路透社上週報導，中國AI代理在測試中展現出欺騙使用者、規避限制並隱瞞失敗的能力，呼應了外界對美國先進系統提出的擔憂。Reuters reported last week that Chinese AI agents had shown an ability to deceive users, evade restrictions and conceal failures in tests, echoing concerns raised about advanced US systems.

中國最新的《AI安全治理框架》列出多項風險，包括模型未經授權取得系統權限或外部資源、欺騙評估人員、隱藏能力，以及繞過安全控制。不過，SemiAnalysis指出，該框架並未依模型能力課予強制性義務。China's latest AI Safety Governance Framework identifies risks including models acquiring system permissions or external resources without authorization, deceiving evaluators, concealing capabilities and bypassing safety controls. But it does not impose mandatory duties linked to model capability, according to SemiAnalysis.

報告表示，北京具有約束力的規定主要規範的是應用及其對使用者的影響，而非要求前沿模型開發商依據模型能力進行或公布風險評估。The report said Beijing's binding rules principally govern applications and their effects on users rather than requiring frontier developers to conduct or publish risk assessments based on a model's capabilities.

這家美國研究公司並補充說，中國主要開發商尚未發布任何公開揭露危險能力測試的前沿文字模型；這類測試涵蓋網路、生物及失控風險。The US research firm added that no major Chinese developer had released a frontier text model with publicly disclosed dangerous-capability tests spanning cyber, biological and loss-of-control risks.

該報告未提供美國AI開發商的可比較數據。包括OpenAI、Anthropic和Google DeepMind在內的美國領先企業，已針對部分重大前沿模型發布，公布安全報告、系統卡或模型卡。The report did not provide comparable figures for US AI developers. Leading US companies including OpenAI, Anthropic and Google DeepMind have published safety reports, system cards or model cards for some major frontier-model launches.

（愛德華多・巴普蒂斯塔報導，露易絲・海文斯編輯）(Reporting by Eduardo Baptista, Editing by Louise Heavens)