NVIDIA：AI 安全是工程问题，需覆盖 Agent 栈每一层
原标题：AI Security Is an Engineering Problem — How to Solve It at Every Layer of the Agent Stack
AI 摘要
NVIDIA 发布博客文章，主张 AI 安全是一个工程问题，需要明确的安全需求、可执行的管控、指定负责人和可验证的证据。文章提出安全应覆盖 agent 技术栈的每一层，包括模型、编排框架和运行时环境，并介绍了 NVIDIA OpenShell 开源安全运行时及 Open Secure AI Alliance 合作伙伴（Cisco、JFrog、CrowdStrike、Palo Alto Networks 等）的相关工具。
正文节选
AI security is an engineering problem. That means defined security requirements, enforceable controls, named owners and evidence that protections work. As AI becomes more capable, the industry must accelerate security engineering, broaden access to defensive tools and share what works faster. Technology Changes, Security Fundamentals Endure The internet and cloud computing changed how software operates, while core security responsibilities endured: establish identity, control access, limit expos