Microsoft and Nvidia publish open-weight model risk management guidance
Microsoft and Nvidia on October 16 jointly published a guide titled "Risk Management for Open-Weight Models in Enterprise," releasing the risk classification material on GitHub and inviting outside suggestions. The guide groups risks into five categories — cybersecurity, data protection, content, human-AI interaction and bias — covering 30 risk types, and offers recommendations on assessment, safeguards and continuous monitoring. It stresses that whether weights are open is not the sole basis for judging risk, since enterprises must weigh model characteristics, intended use, deployment method and protections; it cites coding uses that may produce vulnerable or malicious code and agent uses that raise improper tool-call and excessive-permission concerns.
The guidance gives enterprises deploying open-weight models, whether self-hosted or via managed services, a checklist to assess against, and warns that fine-tuning, quantization or merging models may weaken safeguards and requires retesting.
微軟與Nvidia於10月16日共同發布開放權重AI模型風險管理指引,文件名稱為Risk Management for Open-Weight Models in Enterprise。這份指引將相關風險歸納為網路安全(Cybersecurity)、資料保護(Data Protection)、內容風險(Content)、人類與AI互動(Human-AI Interaction)及偏見(Bias)等5大類、30種類型,並提出風險評估、防護措施及持續監控等管理建議。
這份指引強調,模型權重是否開放並非判斷AI系統風險高低的唯一依據,企業仍需根據模型特性、實際用途、部署方式及防護措施評估風險。例如,企業使用開放權重AI模型協助程式開發,可能產生含有安全漏洞或惡意功能的程式碼;若將模型用於AI代理,則需注意模型可能產生不當的工具呼叫,以及代理存取權限過大等問題。此外,還需注意訓練資料或個人資料外洩、模型遭竊,以及多租戶環境的資料隔離失效等問題。由於開放權重模型不一定附帶完整的安全防護功能,企業自行部署時,可能需要建立身分驗證、存取控制及輸入與輸出過濾等措施;即使使用託管服務,也應確認服務供應商提供哪些防護措施,以及哪些措施仍需由企業自行負責。