ADR

ADR, short for Agentic AI Detection and Response, is an open-source security system pointed at the AI agents a company is already running: the coding agents employees use, such as Cursor, Claude Code and Codex, and customer-facing ones such as AI support. A sensor gathers agent intent, tool calls and execution traces on macOS, Linux and Windows, and detection runs in two tiers, with high-recall triage first and deeper reasoning only on the sessions that look suspicious.

Whether those defenses hold is tested by ADR-Bench, which reproduces the 17 known agent attack techniques across 303 tasks and 133 MCP servers. Uber runs the system in production and the accompanying paper was accepted to MLSys 2026. The prevention component is not part of this open-source release.