Replit publishes Agent Habitat security architecture for AI agents
Replit published the Agent Habitat Architecture white paper, proposing a controlled environment where AI agents interact with people, knowledge bases, tools, and business systems under explicit rules. The framework combines credential isolation, policy-checked network access, human approval for irreversible high-risk actions, and an 11-dimension maturity model spanning Levels 0 through 3. These capabilities and security guarantees are Replit vendor claims reported by DoNews and have not been independently validated here.
The white paper gives enterprises a concrete reference architecture and assessment scale for securing agents that can call tools and execute code, though its practical effectiveness remains unverified.
近日，在线编程平台 Replit 对外发布《Agent Habitat Architecture》白皮书，推出一套面向 AI 智能体的安全参考架构与成熟度评估体系，旨在解决当下 AI Agent 权限失控、提示注入、凭证泄露、越权逃逸等安全风险，为企业部署具备工具调用、代码执行能力的 AI 智能体提供可落地的安全基准。
随着 AI Agent 能力快速迭代，智能体已经能够编写代码、调用第三方工具、访问企业内部业务系统。但当前多数 Agent 运行环境缺少原生安全边界：智能体可获取密钥、读取私有数据，容易被网页、文档中的恶意指令诱导，发生数据外泄、越权删改业务资源等安全事故。白皮书提出 “Agent Habitat（智能体栖息地）” 概念：即企业定义的受控运行环境，让 AI 智能体在明确规则下与人、知识库、工具、业务系统交互，在释放 Agent 生产力的同时建立完整安全管控边界。
核心设计思想：Agent 运行过程中权限只能收缩，不能扩大；凭证永远对智能体不可见；所有网络访问必经策略校验；不可逆高危操作必须经过人工审批。成熟度等级 1 抵御意外故障，等级 2 防御提示注入攻击，等级 3 应对智能体主动尝试逃逸的极端场景。
#Replit#Agent Habitat Architecture#Identity bootstrapper#Harness#Durable executor#Token service#Sandbox#Broker#NVIDIA Open Agent Safety Platform#Blueprint Alliance