uber / ADR

项目摘要

ADR is an enterprise security system for AI agents that provides discovery, observability, benchmarking, threat detection, and prevention capabilities. Deployed in production at Uber and backed by an MLSys 2026 paper, it secures employee-facing AI tools like Cursor, Claude Code, and Codex, as well as customer-facing agents.

为什么值得关注

This project addresses a critical and underserved need — treating AI agents as security endpoints requiring monitoring, detection, and response, analogous to EDR/XDR for traditional endpoints. Its production deployment at Uber and academic paper validation give it strong credibility, while its multi-component architecture covers the full security lifecycle for AI agents.

优势

• Production-proven at a major tech company (Uber) with academic validation at MLSys 2026
• Comprehensive five-layer security architecture covering discovery, observability, benchmarking, detection, and prevention
• Broad coverage of AI coding tools (Cursor, Claude Code, Codex, GitHub Copilot CLI, DeepSeek Harness) and MCP servers
• Includes ADR-Bench with 300+ tasks and coverage of all 17 agent attack techniques
• Two-tier detection architecture combining triage with agentic reasoning for efficient threat detection

局限性

• ADR Prevention component is not yet open-sourced, limiting full end-to-end security coverage
• No tests or Docker support, reducing reproducibility and ease of deployment
• No examples provided, which may slow onboarding for new users
• ADR Explorer (offline red teaming engine) is also not included in the open-source release

使用场景

• Securing enterprise AI coding assistant deployments across development teams
• Monitoring and detecting prompt injection and other AI-specific threats in agent workflows
• Benchmarking and evaluating agent security posture against known attack techniques
• Discovering and inventorying unauthorized AI tool usage across employee endpoints
• Observability and auditing of AI agent actions for compliance and incident response