Edera 公司提出了一种名为“受限权威”(Bounded Authority)的安全模型,旨在解决 AI 代理在访问敏感资源时面临的凭证泄露风险 1。该模型基于对象能力(OCAP)理论,主张不向工作负载直接提供通用凭证或环境权限,而是授予仅能执行特定操作的窄范围能力引用 1。通过这种机制,即使 AI 代理被攻破,其可执行的恶意操作也被严格限制在既定能力范围内,从而提升生产环境中智能体基础设施的安全性 1。
传统安全模式假设工作负载拥有权限后加以约束,而对象能力模型要求工作负载在未持有能力前无任何权限 1。Edera 提出的“受限权威”是故意围绕特定资源、操作或身份进行约束的对象能力 1。对于 LLM 代理而言,能力模型比基于行为的策略更有效,因为系统只需强制权限边界,无需预测或拦截恶意意图(如提示注入) 1。
在具体实现上,核心概念包括 capcurl(HTTP API 的受限权威实现)和 capsudo(将特定命令作为能力暴露,而非赋予 sudo 权限) 1。该架构强调身份(Identity)与权限(Authority)分离:使用 SPIFFE/SPIRE 建立身份,使用能力建立具体的操作权限 1。此外,Edera 计划通过 Project Lunchbox 等项目,将能力作为 Kubernetes 原生的一等基础设施工具进行管理 1。
Edera has introduced a security framework called "Bounded Authority," designed to mitigate credential leakage risks for AI agents accessing sensitive resources 1. Grounded in object capability (OCAP) theory, this model rejects the traditional approach where workloads are granted broad permissions and then constrained 1. Instead, it advocates providing only narrow-scope capability references that allow specific operations, ensuring that even if an agent is compromised, its ability to execute malicious actions remains strictly limited within those predefined boundaries 1.
The core philosophy of Bounded Authority is that workloads should possess no privileges until they explicitly hold a capability, contrasting with legacy models that assume permissions exist and must be restricted 1. This separation of identity from authority is achieved by using SPIFFE/SPIRE for establishing identity and capabilities for defining operational rights 1. Key implementations include capcurl, which serves as a bounded authority interface for HTTP APIs, and capsudo, which exposes specific commands as capabilities rather than granting general sudo access 1. For Large Language Model (LLM) agents, this capability-based model proves more effective than behavior-based policies because the system only needs to enforce permission boundaries rather than attempting to predict or intercept malicious intents such as prompt injection 1. Looking ahead, Edera plans to manage these capabilities as native, first-class infrastructure tools within Kubernetes through initiatives like Project Lunchbox 1.
评论
还没有评论,欢迎留下第一条。