📄 中文摘要
在工程化自主 AI 代理时，从本地开发到生产部署的过渡带来了重要的架构挑战。在隔离环境中，代理能够成功接收提示、制定计划、触发工具序列并执行任务。然而，当部署到多租户生产环境时，出现了一种危险的漏洞：一旦代理开始链接操作，用户身份就会消失。在复杂的 orchestration 工作流程的第三步，尤其是在代理执行涉及实际资金移动或数据删除的 API 调用之前，系统通常只能看到来自一个通用的、全能的服务请求，这使得身份验证和安全性面临重大风险。
保护 AI 代理工作流程：防止多步骤链中的身份崩溃
📄 English Summary
Securing AI Agent Workflows: Preventing Identity Collapse in Multi-Step Chains
Transitioning from local development to production deployment of autonomous AI agents presents a significant architectural challenge. In isolated environments, agents can successfully take prompts, formulate plans, trigger sequences of tools, and execute tasks. However, in a multi-tenant production environment, a critical vulnerability arises: once agents start chaining actions, user identity dissolves. By the third step of a complex orchestration workflow—especially just before executing an API call involving actual money movement or data deletion—the system often only recognizes a request from a generic, omnipotent service, posing substantial risks to authentication and security.