AI News AI资讯 3d ago Updated 3d ago 更新于 3天前 43

Meet SAM (Sovereign Agent Mesh): A Zero-Config, Zero-Trust P2P Network for AI Agents 认识 SAM(主权代理网格):一种面向 AI 代理的零配置、零信任 P2P 网络

SAM (Sovereign Agent Mesh) is an Apache-2.0 P2P networking project enabling autonomous AI agents to share tools across heterogeneous environments without exposing internal APIs to the public internet It uses a three-binary architecture: sam-control-plane (identity/policy), sam-router (libp2p transport), and sam-node (MCP client), with OIDC-to-Biscuit token translation for offline authorization The system enforces strict default-deny access control, where even service discovery requires explicit SAM (Sovereign Agent Mesh) 是Apache-2.0开源的P2P网络项目,专为跨多网络边界的自主AI代理工具共享设计,非Segment Anything模型 采用零配置、零信任架构,基于libp2p和Model Context Protocol实现节点自动发现和加密授权 核心架构包含三个组件:控制平面(身份/策略)、路由器(P2P传输)、节点(MCP接口),支持OIDC到Biscuit令牌的离线授权转换 严格的默认拒绝安全策略,所有访问需显式授权,包括服务发现目录 适用于金融、医疗、政府、国防等受监管行业,生产环境需自建控制平面

62
Hot 热度
65
Quality 质量
60
Impact 影响力

Analysis 深度分析

TL;DR

  • SAM (Sovereign Agent Mesh) is an Apache-2.0 P2P networking project enabling autonomous AI agents to share tools across heterogeneous environments without exposing internal APIs to the public internet
  • It uses a three-binary architecture: sam-control-plane (identity/policy), sam-router (libp2p transport), and sam-node (MCP client), with OIDC-to-Biscuit token translation for offline authorization
  • The system enforces strict default-deny access control, where even service discovery requires explicit capability grants, and supports local attenuation for time-based or role-based restrictions
  • Production deployment requires self-hosting the control plane ("DIY Mode"), as the public testnet remains a beta; the project ships Go binaries, Docker images, Helm charts, and Android/iOS support
  • Key use cases include cross-cloud MCP tool sharing, hybrid on-prem-to-cloud agent calls, sandboxed agents with credential injection via the Secure Outbound Gateway, and pooled warm workers for batch workloads

Why It Matters

SAM addresses a critical infrastructure gap in the agentic AI ecosystem: secure, zero-trust inter-agent communication across network boundaries without requiring agents to operate in exposed or untrusted environments. For AI practitioners deploying agents across cloud, on-prem, and edge devices, SAM provides a production-grade networking layer that eliminates the need to publish internal tools to the internet while maintaining cryptographic authorization and offline verification capabilities.

Technical Details

  • Architecture: Three core binaries — sam-control-plane handles identity registration, token issuance, and policy distribution; sam-router provides libp2p bootstrap points and GossipSub routing overlays; sam-node acts as the P2P client with mesh transport, self-healing connectivity, and a local MCP HTTP interface on port 8080
  • Identity & Authorization: OIDC JWTs are verified by the control plane and translated into Datalog facts sealed into Biscuit tokens (e.g., sub becomes user(...), groups become group(...), peer ID binds as client_peer_id(...)). Nodes authorize offline by evaluating tokens against local rules without calling home, using a two-stage pipeline with ban/revocation cache checks and dual Biscuit authorizer passes
  • Access Control: Strict default-deny enforcement where every capability (including system://sam.catalog discovery) requires explicit granted_service_exact(...) facts. Services follow a type://name convention with wildcard support (mcp://*, mcp://build-runner.*). Local attenuation can deny writes after hours or block contractors, but cannot bypass control-plane check_if constraints
  • Security & Egress: The Secure Outbound Gateway (sam-box/nano-init) addresses agent security gaps by running as PID 1 in sandboxes, setting proxy env vars, and using LD_PRELOAD to intercept C connect() syscalls on ports 80/443. Traffic reaches sam-box via Unix domain socket, where Biscuit verification and credential injection from secrets.yaml occur before HTTPS upgrade — the agent sandbox never holds credentials
  • Deployment & Integration: Ships Go binaries, install script, ghcr.io Docker images, Helm chart (charts/sam-mesh), production Kubernetes guide, and Android/iOS support. Public testnet at bananas.sam-mesh.dev. Exposes standard MCP tools (discover_remote_services, find_remote_tools, call_remote_tool) with guides for Gemini, Claude Code, Claude Desktop, Google Antigravity, and OpenClaw. The sam-node skill install command writes a SKILL.md for agent self-enrollment while keeping human-controlled login

Industry Insight

  • Organizations running AI agents across multiple network boundaries (cloud, on-prem, edge) should evaluate SAM as a networking layer that eliminates the security trade-off between agent interoperability and internet exposure — particularly regulated sectors like financial services, healthcare, and defense that cannot publish internal tools publicly
  • The OIDC-to-Biscuit translation pattern offers a reusable blueprint for zero-trust agent authorization systems, enabling offline verification while maintaining centralized policy control — this approach could influence future standards for agent-to-agent communication security
  • The public mesh remains in beta; enterprises should plan for self-hosted control plane deployments to achieve full data and policy sovereignty, and should monitor the project's evolution as agent networking becomes a critical infrastructure concern in multi-agent systems

TL;DR

  • SAM (Sovereign Agent Mesh) 是Apache-2.0开源的P2P网络项目,专为跨多网络边界的自主AI代理工具共享设计,非Segment Anything模型
  • 采用零配置、零信任架构,基于libp2p和Model Context Protocol实现节点自动发现和加密授权
  • 核心架构包含三个组件:控制平面(身份/策略)、路由器(P2P传输)、节点(MCP接口),支持OIDC到Biscuit令牌的离线授权转换
  • 严格的默认拒绝安全策略,所有访问需显式授权,包括服务发现目录
  • 适用于金融、医疗、政府、国防等受监管行业,生产环境需自建控制平面

为什么值得看

SAM解决了AI代理跨云、数据中心和边缘设备运行时共享工具的安全痛点,为多网络边界场景提供了零信任P2P解决方案,对受监管行业的AI部署具有重要参考价值。

技术解析

  • 架构设计:三个核心二进制组件——sam-control-plane负责身份注册、令牌签发和策略分发;sam-router基于libp2p提供引导节点和GossipSub路由覆盖;sam-node作为P2P客户端提供网格传输、自愈连接和本地MCP HTTP接口
  • 身份与授权:控制平面验证OIDC JWT后将其转换为Datalog事实并封装为Biscuit令牌,节点可离线验证令牌而无需回连控制平面,实现去中心化授权
  • 安全策略:采用绝对默认拒绝机制,所有访问需显式授权,服务发现目录system://sam.catalog也需单独授权,服务命名遵循type://name约定并支持通配符
  • 请求处理:每个请求经过两阶段管道——第一阶段检查封禁和撤销缓存,第二阶段执行两次Biscuit授权验证,并通过连接对等方ID匹配防止重放攻击
  • 出口控制:sam-box和nano-init实现安全出站网关,通过LD_PRELOAD拦截C connect()系统调用,在Unix域套接字上验证Biscuit令牌并注入凭据,代理沙箱无需持有密钥

行业启示

  • 随着AI代理从单VPC扩展到多云、数据中心和边缘设备的混合部署,零信任P2P网络将成为跨边界工具共享的基础设施标准
  • 受监管行业(金融、医疗、政府、国防)对AI代理的安全需求将推动类似SAM的零配置、离线授权方案成为刚需
  • 企业应优先在跨网络边界的代理场景中部署此类方案,单一VPC内的初创公司收益有限,价值体现在多云+本地+边缘的混合架构中

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。

Agent Agent Open Source 开源 Security 安全 LLM 大模型 Deployment 部署