uber/ADR

ADR secures enterprise AI agents through observability, security benchmarking, and threat detection. Deployed at Uber.

What it solves

ADR provides an enterprise security system designed to protect organizations from the risks associated with AI agents. It addresses the need to monitor and secure both employee-facing tools (like Cursor, GitHub Copilot, and Claude Code) and customer-facing support agents, preventing unsafe actions and detecting threats in real-time.

How it works

ADR uses a multi-layered approach to security:

  • Discovery: It inventories all AI applications, CLI agents, IDE extensions, and model runtimes on employee endpoints to find unsanctioned tools.
  • Observability: It captures agent intent, tool usage, and execution traces across various operating systems (macOS, Linux, Windows) to understand agent behavior.
  • Detection: It employs a two-tier architecture that combines high-recall triage with deep agentic reasoning to identify risky behavior efficiently.
  • Benchmarking: It uses ADR-Bench, which contains over 300 tasks and 134 MCP servers, to test agent security under realistic enterprise conditions.

Who it’s for

This system is intended for enterprise security teams and organizations deploying AI agents at scale who need to maintain visibility and control over agentic AI activity.

Highlights

  • Production Proven: Deployed in production at Uber.
  • Broad Compatibility: Supports a wide range of AI coding tools and OS platforms.
  • Harnesses MCP: Integrates with Model Context Protocol (MCP) servers for benchmarking and detection.
  • Comprehensive Benchmarking: Covers all 17 known agent attack techniques via ADR-Bench.

Related

  • Project
    secureagentics/Adrian

    An open-source runtime security engine for AI agents that monitors both actions and reasoning traces to detect and block malicious or out-of-remit behavior in real time.

  • Project
    cordum-io/cordum

    An agent control plane for governance, safety, and trust that provides a deterministic layer to define, enforce, and audit the behavior of autonomous AI agents.

  • Project
    rogue-security/rogue

    An AI agent evaluator and red teaming platform that stress-tests agents against business policies and adversarial attacks to identify security vulnerabilities.

  • Dispatch
    Anthropic Trustworthy Agents Framework and Implementation

    Anthropic outlines a multi-layered approach to building trustworthy AI agents by combining model training, system harnesses, tool permissions, and open standards to mitigate risks like prompt injection and unintended autonomy.

  • Project
    cisco-ai-defense/skill-scanner

    Skill Scanner is an open‑source security scanner for AI‑agent skill packages (e.g., OpenAI Codex, Cursor, Claude). It combines signature‑based YARA/YAML rules, AST‑based data‑flow analysis, optional LLM‑as‑a‑judge, and a CEL decision layer to detect prompt‑injection, data‑exfiltration, and malicious code patterns. Install via pip, run from the CLI or Python SDK, and integrate with CI/CD via GitHub Actions or a pre‑commit hook. Findings are reported with severity levels; a clean scan is best‑effort only and does not guarantee safety.