ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
TopicDisconfirmAI EditorialUpdated 2026-10-18 07:43

智能体AI安全治理

4 steps3 pieces5 Sources

随着AI智能体从聊天机器人演变为能够自主规划、调用工具并跨系统操作的实体，其安全边界已从模型输出扩展至执行链、身份认证及网络出口。当前治理核心在于解决传统IAM无法应对的“非人类身份”风险：如权限动态累积、缺乏归因审计及未受控的网络连接。

【背景】微软Foundry团队指出，仅靠工具列表不等于网络边界，需通过RAI策略定义出站目的地规则；而清华DeepKernel Lab则提出AgentKernel概念，主张在操作系统层面建立不可绕过的强制信任边界，涵盖身份、感知、认知与执行四大支柱。

争议焦点在于治理层级：是依赖应用层中间件（如现有合规框架）还是下沉至OS内核或区块链共享信任层？前者易被绕过，后者实施复杂度高但具备结构性优势。

How it got here

  1. 2026-08-29

    清华大学DeepKernel Lab发布论文《AgentKernel》，提出首个信任原生智能体操作系统，主张将安全作为第一类设计约束，通过内核级强制服务解决提示注入和内存中毒问题。

    arxiv.org ↗
  2. 2026-10-14

    IBM举办网络研讨会，强调智能体身份管理的四个关键缺口：无治理的持久访问、未经审查的权限提升、缺乏归因以及碎片化控制，推荐采用Vault和Verify实现零信任动态凭证。

    ibm.com ↗
  3. 2026-10-18

    微软Foundry博客演示如何通过RAI策略为托管智能体设置可测试的网络出口边界，区分审计与强制模式，明确工具列表不等于网络边界。

    devblogs.microsoft.com ↗
  4. 2026-10-18

    前Google Cloud CISO Phil Venables在播客中探讨CISO向“首席数字风险官”转型，聚焦企业级AI智能体的控制平面及新兴安全合规框架。

    youtube.com ↗

What is still unsettled

  • 操作系统级安全架构（如AgentKernel）在实际生产环境中如何平衡性能开销与安全强制性？
  • 在开放智能体网络中，基于区块链的信任机制能否有效解决跨组织身份验证与价值结算的可扩展性问题？
  • 现有的应用层治理工具（如Microsoft AGT）是否足以应对日益复杂的代理间协作攻击，还是需要更底层的内核支持？

Sources

Sign in / Sign up
ReadingResearchRadarInvestment framework
  • arxiv.orgAgentKernel论文↗
  • devblogs.microsoft.com微软Foundry网络出口控制指南↗
  • ibm.comIBM智能体身份管理方案↗
  • github.com微软AGT威胁模型文档↗
  • youtube.comPhil Venables访谈↗

Coverage in this topic

  • Microsoft Foundry Blog: Setting Testable Network Egress Boundaries for Hosted AI Agents2026-10-18News

    Muzz Imam from the Microsoft Foundry team explains why tool lists are not network boundaries and demonstrates how to define, attach, and test outbound destination rules for hosted agents using RAI policies, distinguishing between audit and enforced modes.

  • Former Google Cloud CISO Phil Venables on AI Agent Governance2026-10-18News

    Hosted by Noma Security CISO Diana Kelley, this episode features Phil Venables, former CISO of Google Cloud with nearly two decades at Goldman Sachs. The discussion focuses on how AI expands CISO responsibilities towards becoming a "Chief Digital Risk Officer" and explores enterprise agentic control planes and emerging security frameworks.

  • UiPath CISO Scott Roberts: Identity and Trust Boundaries for AI Agents2026-10-18News

    In this episode of The Blast Radius podcast, Scott Roberts, Chief Information Security Officer at UiPath, joins Diana Kelley, CISO of Noma Security, to discuss securing agentic AI at scale. Roberts draws on his extensive experience from Microsoft and Coinbase Cloud to explore challenges in managing identities, enforcing trust boundaries, and achieving machine-speed incident response.