Microsoft Moves AI Governance From Policy to Runtime Enforcement

微软将AI治理从政策转向运行时执行

作者：Leela Kumili · InfoQ 原文

摘要：微软推出一套AI治理架构，将治理从文档政策转向运行时执行、持续评估、可观测性和审计证据。该架构涵盖九个治理领域和四个功能（政策、控制、可见性、证明），形成持续运营循环。结合Microsoft Foundry等工具，实现运行时控制与评估，并借鉴NIST框架。还涉及代理治理和开源工具包，确保AI应用在投产中的安全与合规。

Microsoft has outlined an AI governance architecture that moves governance from documented policies toward runtime enforcement, continuous evaluation, observability, and audit evidence as organizations deploy AI applications and agents in production.

微软推出了一套AI治理架构，将治理从文档化政策转向运行时执行、持续评估、可观测性和审计证据，以应对组织在生产环境中部署AI应用和代理的需求。

The framework spans nine governance domains and four functions: policy, control, visibility, and proof, addressing the need to verify that governance requirements are enforced and observable during AI system operation.

该框架涵盖九个治理领域和四项功能：政策、控制、可见性和证明，以满足在AI系统运行期间验证治理要求得到执行和可观测的需求。

The architecture treats governance as a continuous operational loop.

该架构将治理视为一个持续运营的循环。

Policies establish requirements and risk classifications, controls translate them into access and runtime rules, observability captures system behavior, and evaluations test quality and safety.

政策确立要求和风险分类，控制措施将其转化为访问和运行时规则，可观测性捕获系统行为，评估测试质量和安全性。

Audit processes then turn operational telemetry into evidence for compliance and incident investigation.

审计流程随后将运营遥测数据转化为合规和事件调查的证据。

Manasa T. Ramalinga, Cloud Solution Architect at Microsoft, described the motivation in a LinkedIn post accompanying the architecture.

微软云解决方案架构师Manasa T. Ramalinga在一篇随架构发布的LinkedIn帖子中描述了动机。

She added that organizations moving AI workloads into production are re-architecting their foundational structures to build safer systems, rather than treating governance as an afterthought.

她补充说，将AI工作负载投入生产的组织正在重新架构其基础结构以构建更安全的系统，而不是将治理视为事后考虑。

Organizations cannot scale what you cannot control

组织无法扩展你无法控制的东西。

Microsoft identifies nine governance domains: policy, data governance, model governance, observability, evaluations, security, identity and access, audit and compliance, and agent governance.

微软确定了九个治理领域：政策、数据治理、模型治理、可观测性、评估、安全、身份与访问、审计与合规、以及代理治理。

Runtime controls can span interactions among users, agents, models, tools, APIs, MCP servers, and enterprise systems.

运行时控制可以涵盖用户、代理、模型、工具、API、MCP服务器和企业系统之间的交互。

The architecture combines Microsoft Foundry with services including Microsoft Purview, Microsoft Entra ID, Defender, and Azure API Management.

该架构将Microsoft Foundry与Microsoft Purview、Microsoft Entra ID、Defender和Azure API Management等服务相结合。

Foundry's AI Gateway provides a runtime boundary for authentication, token limits, quotas, and policy enforcement.

Foundry的AI网关为身份验证、令牌限制、配额和政策执行提供了运行时边界。

Microsoft also documents using the gateway to govern MCP tools, providing centralized authentication, rate limiting, IP restrictions, and audit logging without modifying MCP servers or agent code.

微软还记录了使用该网关来治理MCP工具，提供集中式身份验证、速率限制、IP限制和审计日志记录，而无需修改MCP服务器或代理代码。

Evaluations are positioned both before deployment and in production.

评估被定位在部署之前和生产环境之中。

Microsoft Foundry supports evaluation of AI applications and agents against datasets using built-in and custom evaluators, allowing teams to assess quality and safety before release and monitor production behavior afterward.

Microsoft Foundry支持使用内置和自定义评估器对AI应用和代理进行数据集评估，使团队能够在发布前评估质量和安全性，并在之后监控生产行为。

Anthony Bartolo, Principal Cloud Advocate at Microsoft, described the operational distinction in a LinkedIn post.

微软首席云倡导者Anthony Bartolo在一篇LinkedIn帖子中描述了运营上的区别。

Your AI policy is not governance until production can prove it.

你的AI政策不是治理，直到生产环境能够证明它。

He outlined the loop as policy defining rules, runtime controls enforcing them, observability capturing behavior, evaluations testing quality and safety, and audit converting telemetry into evidence.

他将循环描述为：政策定义规则，运行时控制执行规则，可观测性捕获行为，评估测试质量和安全性，审计将遥测数据转化为证据。

The approach is broader than a Microsoft-specific control plane.

这种方法比微软特有的控制平面更广泛。

The NIST AI Risk Management Framework and Generative AI Profile provide a vendor-neutral framework for managing AI risks across the lifecycle, including governance, measurement, evaluation, and risk mitigation.

NIST AI风险管理框架和生成式AI配置文件提供了一个供应商中立的框架，用于管理整个生命周期的AI风险，包括治理、测量、评估和风险缓解。

Microsoft's architecture maps those governance concerns into concrete platform controls and operational telemetry.

微软的架构将这些治理问题映射到具体的平台控制和运营遥测中。

Agent governance adds controls around agent identity, access, activity, and workflow checkpoints.

代理治理增加了围绕代理身份、访问、活动和工作流检查点的控制。

Microsoft's open-source Agent Governance Toolkit provides runtime security capabilities for autonomous agents, including policy enforcement and interception points.

微软的开源代理治理工具包为自主代理提供了运行时安全能力，包括政策执行和拦截点。

The architecture also describes the Agent Control Specification as a mechanism for checkpoints across agent inputs, model calls, tool execution, and outputs, with higher-impact actions able to require human approval.

该架构还将代理控制规范描述为一种跨代理输入、模型调用、工具执行和输出的检查点机制，高影响力操作可能需要人工批准。

阅读理解

1. Which of the following is NOT one of the four functions in Microsoft's AI governance framework?

2. What does Microsoft's AI Gateway provide for governing MCP tools, according to the article?

3. According to Anthony Bartolo, when does an AI policy become governance?

温故复习 →每日一句 →