AI Practices AI实践 3d ago Updated 3d ago 更新于 3天前 46

How Axonius built secure multi-tenant AI agents on Bedrock AgentCore Axonius如何在Bedrock AgentCore上构建安全的多租户AI代理

Axonius adopted a hybrid multi-tenancy architecture for AI agents on Amazon Bedrock AgentCore, combining a shared runtime with infrastructure-level tenant enforcement at the tool layer via AgentCore Gateway The platform supports three deployment patterns—silo, pool, and bridge—each offering different trade-offs between isolation, cost, and operational complexity Tenant isolation is enforced through JWT-based authentication with custom tenant claims, enabling secure routing of tool calls and data Axonius作为安全智能平台ISV,在保持silo部署模式的同时引入AI代理,解决多租户场景下的安全隔离与运维挑战 文章系统介绍了三种多租户AI代理架构模式:silo(租户专属资源)、pool(共享资源+会话级隔离)、bridge(混合模式) Axonius的核心需求涵盖租户数据隔离、现有身份系统集成、按租户成本追踪、API安全访问、CD生命周期管理及大规模agent可观测性 Amazon Bedrock AgentCore通过microVM隔离和内置JWT授权器,支持在共享runtime中实现基于tenant_id声明的租户级路由 混合方案结合共享runtime的运维简化与AgentCor

62
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Axonius adopted a hybrid multi-tenancy architecture for AI agents on Amazon Bedrock AgentCore, combining a shared runtime with infrastructure-level tenant enforcement at the tool layer via AgentCore Gateway
  • The platform supports three deployment patterns—silo, pool, and bridge—each offering different trade-offs between isolation, cost, and operational complexity
  • Tenant isolation is enforced through JWT-based authentication with custom tenant claims, enabling secure routing of tool calls and data access per customer
  • Key requirements addressed include strict tenant data scoping, identity integration with existing AWS Cognito/EC2 auth modules, per-tenant cost tracking, API integration, CI/CD lifecycle management, and fleet-scale observability
  • The hybrid approach eliminates application-dependent isolation risks while preserving operational simplicity of a single shared runtime deployment

Why It Matters

This case study provides a practical blueprint for ISVs and SaaS providers looking to integrate AI agents into multi-tenant environments without compromising security or isolation guarantees. It demonstrates how AWS Bedrock AgentCore's architecture can reconcile the tension between operational efficiency (shared resources) and enterprise-grade tenant separation (dedicated enforcement), which is a critical concern for any organization handling sensitive customer data at scale.

Technical Details

  • Three multi-tenancy patterns: The silo model uses dedicated agents per tenant with full infrastructure isolation; the pool model shares a single runtime across tenants with session-level isolation via unique session IDs; the bridge/hybrid model combines a shared AgentCore runtime with per-tenant enforcement at the tool layer through AgentCore Gateway
  • JWT-based tenant routing: Tenants authenticate through OAuth 2.0 (e.g., Amazon Cognito), with JWTs carrying a custom:tenant_id claim. AgentCore's built-in JWT authorizer validates tokens via a configured discovery URL, and agent code routes tool calls and data access based on the tenant claim
  • MicroVM-level isolation: AgentCore runtime allocates a dedicated microVM for each session, providing structural separation while serving multiple tenants from a single runtime instance
  • Integration requirements: Axonius maintained its existing silo deployment model (dedicated Amazon VPC per customer with ALB, NLB, databases, and EC2 compute), requiring the agent identity flow to integrate seamlessly with its existing authentication and authorization module without disruption
  • Operational concerns addressed: Per-tenant cost tracking (critical since model invocation drives most agentic costs), observability for large agent fleets with tracing and alerting, and inclusion of agentic workloads into existing silo Continuous Delivery pipelines

Industry Insight

  • ISVs should carefully evaluate the hybrid (bridge) model when balancing onboarding speed against security requirements—it enables rapid customer onboarding without infrastructure provisioning while still enforcing tenant boundaries at the infrastructure layer, not just in application code
  • Per-tenant cost tracking must be baked into agentic architecture from the start, as model invocation costs can escalate unpredictably and directly impact pricing strategy for AI-powered SaaS offerings
  • Platform engineers should prioritize observability and lifecycle management early in agent deployment; managing large fleets of per-tenant agents requires robust tracing, alerting, and CI/CD integration to avoid operational blind spots as the tenant base scales

TL;DR

  • Axonius作为安全智能平台ISV,在保持silo部署模式的同时引入AI代理,解决多租户场景下的安全隔离与运维挑战
  • 文章系统介绍了三种多租户AI代理架构模式:silo(租户专属资源)、pool(共享资源+会话级隔离)、bridge(混合模式)
  • Axonius的核心需求涵盖租户数据隔离、现有身份系统集成、按租户成本追踪、API安全访问、CD生命周期管理及大规模agent可观测性
  • Amazon Bedrock AgentCore通过microVM隔离和内置JWT授权器,支持在共享runtime中实现基于tenant_id声明的租户级路由
  • 混合方案结合共享runtime的运维简化与AgentCore Gateway的工具层租户边界强制,提供基础设施级隔离保障

为什么值得看

本文面向SaaS提供商和平台工程师,系统梳理了ISV在引入AI代理时的多租户架构选型路径,提供了从理论模式到落地实现的完整参考。Axonius作为真实案例,展示了如何在保持现有silo部署约束下平滑集成agentic workload,对同类ISV具有直接借鉴价值。

技术解析

  • 三种多租户架构模式:Silo模式为每个租户分配独立agent资源,隔离性最强但运维成本高;Pool模式由单一agent服务多租户,通过唯一会话ID隔离用户会话,运维简化但依赖应用层隔离;Bridge模式混合使用两种策略,例如agent独立部署但共享Amazon Bedrock Knowledge Bases(托管RAG服务)。
  • Axonius的Silo部署约束:每个客户工作负载位于独立Amazon VPC内,包含ALB、NLB、数据库和通用计算基础设施。Axonius选择维持现有租户管理方法论,不改变silo部署模型。
  • 六大关键需求:租户隔离要求agent仅能访问所属客户数据;身份集成需与现有EC2上的认证授权模块无缝对接;成本追踪聚焦模型调用费用以支撑定价策略;服务集成确保agent安全访问租户API;生命周期管理需融入现有silo CD工作流;可观测性需支持大规模agent集群的监控、告警和调试追踪。
  • Pool模式技术实现:AgentCore runtime为每个会话分配独立microVM,通过OAuth 2.0身份提供商(如Amazon Cognito)认证,JWT携带custom:tenant_id声明。runtime内置JWT授权器通过发现URL获取公钥验证令牌签发者,agent代码读取声明后路由工具调用和数据访问至正确租户环境。
  • 混合架构方案:结合共享runtime的运维简便性与AgentCore Gateway的工具层租户边界强制,租户共享单一runtime,但所有出站工具调用均通过Gateway执行,在工具代码执行前强制实施租户隔离。

行业启示

  • ISV引入AI代理时需根据数据敏感度、定制化和运维能力权衡架构模式:高安全要求场景适合silo或bridge,快速迭代和成本优化场景可考虑pool模式。
  • 多租户AI代理架构已成为SaaS厂商的必选项,AWS Bedrock AgentCore等平台正提供标准化的租户隔离、身份集成和可观测性能力,降低ISV的agent部署门槛。
  • 成本追踪与可观测性是多租户agent部署的关键挑战,建议建立按租户的模型调用成本计量机制,并选择支持大规模agent集群监控告警的平台能力。

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 Security 安全 Deployment 部署 LLM 大模型