LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
← Voltar para a leitura
TópicoEditorial DisconfirmAIAtualizado 2026-11-03 00:02

AI智能体沙箱隔离

5 etapas1 peças10 Fontes

【背景】随着AI编程代理普及,其执行不可信代码带来的安全风险日益凸显。Pillar Security研究员Eilon Cohen与Ariel Fogel在2026年2月的分析指出,容器、gVisor、microVM和内核能力四种主流隔离层级各有失败模式,核心结论是:隔离技术保护宿主机免受沙箱侵害,却无法保护沙箱自身免受内部挂载内容(如凭据)的泄露。

以Cursor为例,CVE-2026-22708显示静态白名单在环境被污染后失效,而完整文件系统读取权限导致凭据可通过标准输出外泄。相比之下,Claude Code通过Seatbelt/Bubblewrap实现双隔离,限制文件访问至当前工作目录并路由网络流量,未报告类似漏洞。OWASP指南强调最小权限原则,建议将工具授权逻辑置于代理上下文之外。学术界则探索更轻量方案,如Sandlock利用Linux非特权原语实现毫秒级启动,SideKernel针对macOS提供易用microVM沙箱。云厂商如腾讯云AGSX和UCloud Sandbox已将此类隔离作为基础设施服务,支持毫秒级启动与多形态沙箱。争议焦点在于:在性能开销与安全边界之间,何种隔离强度能真正阻断“沙箱内”的数据外泄路径?

Como chegamos aqui

  1. 2025-10-20

    Anthropic发布Claude Code沙箱功能,基于OS级原语实现文件系统与网络双重隔离,减少84%权限提示。

    anthropic.com ↗
  2. 2026-02-25

    Pillar Security发布SAIL 2.0框架分析,指出Cursor等代理存在凭据外泄风险,强调隔离层级无法解决沙箱内部挂载问题。

    pillar.security ↗
  3. 2026-03-23

    Sandlock论文在ASPLOS 2026研讨会发表,提出利用Landlock/seccomp实现无特权、低延迟的AI代理进程沙箱。

    arxiv.org ↗
  4. 2026-08-14

    Solo.io博客指出OpenAI、Anthropic等公司近期发生代理逃逸事件,强调硬件隔离(microVM)优于软件隔离。

    solo.io ↗
  5. 2026-09-30

    腾讯云开发者社区文章详解Agent沙箱定义,介绍AGSX服务支持浏览器/代码沙箱及毫秒级启动特性。

    developer.cloud.tencent.com ↗
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

O que ainda está indefinido

  • 如何在不牺牲开发体验的前提下,有效防止沙箱内部敏感凭据通过标准输出或侧信道外泄?
  • 对于本地开发场景,轻量级内核能力隔离(如Landlock)与重型microVM隔离的安全性与性能权衡点在哪里?

Fontes

  • pillar.securityPillar Security Blog↗
  • cheatsheetseries.owasp.orgOWASP AI Agent Security Cheat Sheet↗
  • docs.cloud.google.comGKE Sandbox Documentation↗
  • anthropic.comAnthropic Research: Claude Code Sandboxing↗
  • solo.ioSolo.io: What Is an Agent Sandbox?↗
  • arxiv.orgArXiv: Sandlock Paper↗
  • arxiv.orgArXiv: SideKernel Paper↗
  • developer.cloud.tencent.comTencent Cloud Developer Community Article↗
  • cloud.tencent.comTencent Cloud AGSX Product Page↗
  • github.comUCloud Agent Sandbox GitHub↗

Cobertura neste tópico

  • Pillar Security Researchers: AI Coding Agent Sandboxes Isolate the Host, Not the Sandbox Itself2026-11-02Notícias

    In an analysis published in February 2026, Pillar Security researchers Eilon Cohen and Ariel Fogel noted that while container, gVisor, microVM, and kernel capabilities represent four isolation tiers with distinct failure modes, security is ultimately determined by what is mounted inside the sandbox rather than the strength of the isolation. Using Cursor's CVE-2026-22708 as an example, they demonstrated that static allowlists fail once the environment is poisoned, and granting full filesystem read access allows credentials to be exfiltrated via standard output, even if write permissions are restricted.