AI Security AI安全 4h ago Updated 2h ago 更新于 2小时前 49

Securing Claude Code: The New Compliance API, Local Visibility, and Identity Governance 保护 Claude Code:新的合规 API、本地可见性与身份治理

Anthropic introduced new Compliance API endpoints (as of August 11, 2026) providing visibility into local Claude Code sessions, including session metadata and full transcripts of model interactions Local AI agents account for 68.6% of AI agents discovered in customer environments, inheriting employee credentials, network position, and permissions Claude Code operates as a hybrid system where the LLM (brain) runs in Anthropic's cloud while the harness (hands) executes commands locally on endpoint Anthropic于2026年8月11日推出Claude Code本地会话合规API端点,填补了端点AI代理可观测性空白 本地AI代理已占客户环境中68.6%,继承员工凭证、网络位置和权限,安全治理压力剧增 三层治理架构:托管设置(策略基线)+ 合规API(会话转录)+ 端点遥测(进程/文件/配置) 传统SaaS集中管控模式失效,LLM在云端、执行在端点的混合架构要求安全模型重构 82%的企业发现过安全/IT/治理团队不知情的AI代理,可见性缺口严重

70
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Anthropic introduced new Compliance API endpoints (as of August 11, 2026) providing visibility into local Claude Code sessions, including session metadata and full transcripts of model interactions
  • Local AI agents account for 68.6% of AI agents discovered in customer environments, inheriting employee credentials, network position, and permissions
  • Claude Code operates as a hybrid system where the LLM (brain) runs in Anthropic's cloud while the harness (hands) executes commands locally on endpoints, creating a governance gap
  • Three-layer governance model is needed: managed settings (policy baseline), Compliance API (activity logging), and endpoint telemetry (local execution context)
  • Static allow/deny policies and activity logs alone are insufficient; logs cannot determine whether an agent's access is legitimate, requiring integration of identity, intent, and credentials

Why It Matters

This represents a fundamental shift in AI security as agents move from browser-based interfaces to local endpoints, challenging traditional shared-responsibility models and centralized monitoring approaches. The article exposes a critical governance gap: organizations can now see what local AI agents do, but cannot easily determine if those actions are authorized or legitimate, requiring new security frameworks that combine cloud-side telemetry with endpoint-side context.

Technical Details

  • Compliance API Endpoints: Three new endpoints introduced on August 11, 2026 — GET /v1/compliance/apps/sessions/local (session metadata list), GET /v1/compliance/apps/sessions/local/{session_id} (individual session metadata), and GET /v1/compliance/apps/sessions/local/{session_id}/messages (full transcript)
  • Managed Settings: JSON files on Mac/Linux and registry records on Windows that enforce organizational policy baselines, with precedence over global, project, and user settings; supports allow/deny lists for MCP servers, regex patterns for bash commands, and skill command restrictions
  • Three-Layer Governance Architecture: Layer 1 (Managed Settings for policy baseline), Layer 2 (Compliance API for cloud-side transcript visibility), Layer 3 (Endpoint telemetry for local process/file/configuration data that cloud services cannot observe)
  • Transcript Block Types: Logs capture three block types — text, tool_use, and tool_result — covering user prompts, bash commands, file reads/writes, and MCP commands; the model holds no server-side state, so the harness resends full context each turn
  • Survey Data: Token Security found local agents represent 68.6% of AI agents in customer environments; Cloud Security Alliance survey of 418 IT/security professionals showed 68% rated visibility as high while 82% discovered unknown agents

Industry Insight

  • Organizations must adopt a hybrid governance strategy combining Anthropic's Compliance API with endpoint detection and response (EDR) tools, as neither alone provides complete visibility — EDR offers evidence but cannot connect activity to owner intent and credentials
  • The traditional SaaS centralized monitoring model is insufficient for endpoint AI agents; security teams should expect increased administrative burden and invest in identity governance frameworks that can correlate cloud-side transcripts with endpoint-side execution context
  • Static policy enforcement (allow/deny lists, regex patterns) will prove inadequate for the pace of AI-driven workflows; organizations should plan for dynamic, context-aware governance solutions that can assess legitimacy rather than merely logging activity

TL;DR

  • Anthropic于2026年8月11日推出Claude Code本地会话合规API端点,填补了端点AI代理可观测性空白
  • 本地AI代理已占客户环境中68.6%,继承员工凭证、网络位置和权限,安全治理压力剧增
  • 三层治理架构:托管设置(策略基线)+ 合规API(会话转录)+ 端点遥测(进程/文件/配置)
  • 传统SaaS集中管控模式失效,LLM在云端、执行在端点的混合架构要求安全模型重构
  • 82%的企业发现过安全/IT/治理团队不知情的AI代理,可见性缺口严重

为什么值得看

本文揭示了AI从浏览器向端点迁移后的安全治理新范式,对IT和安全团队理解本地AI代理的风险边界、合规工具链及治理架构具有直接指导价值。

技术解析

  • 合规API端点:新增GET /v1/compliance/apps/sessions/local(会话列表)、GET /v1/compliance/apps/sessions/local/{session_id}(会话元数据)、GET /v1/compliance/apps/sessions/local/{session_id}/messages(完整转录),覆盖text、tool_use、tool_result三类日志块,包含用户提示、bash命令、文件读写及MCP调用。
  • 托管设置(Managed Settings):通过JSON文件(Mac/Linux)或注册表(Windows)实施策略基线,优先级高于全局/项目/用户设置,支持MCP服务器黑白名单、bash命令正则过滤、技能禁用等规则,企业版可通过GUI管理,非企业版依赖MDM下发。
  • 端点遥测局限:EDR可捕获进程、文件和配置证据,但无法关联代理活动与所有者、意图、凭证和权限,需与合规API数据结合形成完整治理视图。
  • Harness架构:LLM仅作为无状态"大脑"处理上下文并决策,实际命令执行、第三方认证、MCP连接均由端点上的harness完成,形成"云端大脑+本地手脚"的混合执行模型。

行业启示

  • 端点安全团队需主导AI代理治理:本地代理继承员工凭证和网络位置,传统云端SaaS管控模式失效,安全团队必须将AI代理纳入端点可见性和身份治理体系。
  • 静态策略无法适应AI动态性:黑白名单和正则过滤缺乏上下文感知能力,企业需向基于意图和上下文的动态策略演进,避免过度限制开发效率。
  • 合规API是治理起点而非终点:Anthropic的端点仅覆盖与模型交互的数据,完整治理需叠加端点遥测、身份绑定和权限审计,形成多层数据融合的可观测框架。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Claude Claude Security 安全 Agent Agent Code Generation 代码生成 Policy 政策