ADR: Agentic AI Detection and Response
An enterprise security system, deployed in production at Uber, that discovers, observes, benchmarks and detects risky behavior in employee-facing and customer-facing AI agents.
ADR, Agentic AI Detection and Response, is an enterprise security system for AI agents, deployed in production at Uber and described in a paper accepted to MLSys 2026. It secures agents such as Cursor, Claude Code, Codex, GitHub Copilot CLI and DeepSeek Harness, as well as customer-facing AI support agents, through five capabilities: discovering unsanctioned AI tools on employee endpoints, observing what agents are doing and why, benchmarking how well defenses hold up under realistic conditions, detecting risky behavior, and, in a component not yet open-sourced, preventing unsafe actions before they cause harm.
The open-source release includes ADR Discovery, which inventories installed AI applications, CLI agents, IDE extensions, local model runtimes and MCP servers and flags unknown ones for review; ADR Sensor, which captures agent intent, tool use and execution traces across macOS, Linux and Windows; ADR-Bench, a benchmark of more than three hundred tasks and 134 MCP servers covering all seventeen known agent attack techniques; and ADR Detector, a two-tier detector combining high-recall triage with deeper reasoning for suspicious sessions. It is written in Python and licensed under Apache 2.0.