uber/ADR
ADR secures enterprise AI agents through observability, security benchmarking, and threat detection. Deployed at Uber.
Python · ★ 1,377 · Fork 120 · 在 GitHub 查看
在手机上浏览源码、订阅每日 AI 趋势日报与项目架构深度解读:获取 CopoHub 客户端(iOS & HarmonyOS)→
GitHub Trending 上榜记录
uber/ADR 已在 GitHub Trending 每日榜单上榜 3 天,最高排名第 4,首次上榜 2026-08-04,最近一次上榜 2026-08-06。
| 日期 | 排名 | Star | 当日增长 |
|---|---|---|---|
| 2026-08-06 | #12 | 1,168 | +354 |
| 2026-08-05 | #4 | 1,010 | +354 |
| 2026-08-04 | #4 | 649 | +140 |
同期上榜的项目
- TencentCloud/TencentDB-Agent-Memory – TencentDB Agent Memory is a team-level memory hub for AI Agents — turning conversations, docs, and code into four reusable memory assets (Chat Memory, Skill, LLM-Wiki, Code-Graph) that are governed, shared, and equipped across agents and frameworks.
- addyosmani/agent-skills – Production-grade engineering skills for AI coding agents.
- cloudflare/computer – Give your agent a computer 👾
- mattpocock/skills – Skills for Real Engineers. Straight from my .agents directory.
- goauthentik/authentik – The authentication glue you need.
- huangruiteng/loopx – Long-horizon agent control plane for durable, governed work across Codex, Claude Code, and other harnesses.
uber/ADR 项目深度剖析与商业评估报告
uber/ADR 项目深度剖析与商业评估报告
报告日期: 2026-08-04 分析对象: uber/ADR (Agentic AI Detection and Response) 项目状态: 已生产部署于 Uber,MLSys 2026 论文支撑
---
1. 核心定位与商业落地价值
1.1 解决的根因性痛点
ADR 瞄准的是 LLM Agent 在企业环境中的"不可观测性"与"不可控性"。传统安全工具(如 WAF、DLP、EDR)的检测对象是明确的字节流或进程行为,而 Agent 的威胁面在于 意图层(Intent) 与 工具调用链(Tool-use Chain)。这导致两个核心问题:
语义鸿沟: 传统规则引擎无法判断"读取客户数据库"是否属于越权行为,需要结合对话上下文与任务意图。 非确定性执行: Agent 的每一步行动由概率模型驱动,无法通过静态签名或沙箱预判。
ADR 通过 四层能力闭环(观察→评估→检测→阻断) 将 Agent 行为纳入企业安全治理体系,本质上解决了 "影子 AI" 泛滥带来的合规审计盲区。
在 README 中展示上榜徽章
[](https://github.fq6825.top/repo/uber/ADR)