Uber Open-Sources ADR, Enterprise Security Framework for AI Agents
Key Takeaways
- ▸Uber open-sourced ADR, a comprehensive security monitoring system for enterprise AI agents
- ▸ADR-Bench includes 300+ benchmark tasks covering all 17 known agent attack techniques across 133 MCP servers
- ▸The system provides observability, detection, and prevention capabilities for AI tools like Claude Code, Cursor, and Codex
Summary
Uber has open-sourced ADR (Agentic AI Detection and Response), a comprehensive security framework designed to monitor and protect enterprise AI agents, including popular tools like Claude Code, Cursor, and Codex. The system was developed in production at Uber and recently had its accompanying paper accepted to MLSys 2026.
ADR provides four key capabilities for securing AI agents: observability (tracking agent intent and tool usage across multiple platforms), benchmarking (ADR-Bench includes 300+ tasks and coverage of all 17 known agent attack techniques), detection (using a two-tier architecture combining triage with agentic reasoning), and prevention (though this component is not yet included in the open-source release). The benchmark covers 133 MCP servers and supports macOS, Linux, and Windows environments.
The open-source release includes the ADR Sensor for telemetry collection, ADR-Bench for security evaluation, and the ADR Detector for threat detection, along with comprehensive documentation. The detection engine uses AI agents powered by Claude to identify suspicious agent behavior patterns, making it both practical for enterprises and accessible to the broader security research community.
- The detection engine uses AI agents (powered by Claude) to identify suspicious behavior through a two-tier architecture
Editorial Opinion
Uber's open-sourcing of ADR addresses a critical gap in enterprise AI security—as organizations increasingly deploy AI agents in production, the ability to monitor and secure them becomes essential. By releasing comprehensive benchmarks and detection tools, Uber provides the industry with both a defensive baseline and a research foundation for improving AI agent security. This move reflects the maturation of the AI agent ecosystem and sets an important precedent for security-first infrastructure in enterprise AI deployments.
