OSS
Uber Open-Sources ADR — an Agent Security Platform With a 303-Task Benchmark Across 133 MCP Servers and 17 Attack Techniques
Uber released ADR (Agentic AI Detection and Response), a production-deployed system for monitoring the AI agents employees actually run — Cursor, Claude Code, and customer-support agents. It ships three parts: a sensor that normalizes telemetry from coding tools on macOS/Linux/Windows, ADR-Bench with 303 tasks spanning 133 MCP servers and 17 documented agent attack techniques, and a dual-agent detector pairing high-recall triage with reasoning-based analysis of suspicious sessions. The accompanying paper was accepted to MLSys 2026; the repo is at 519 stars with +140 today, making it the first major-employer agent-security telemetry stack with a public benchmark attached.
↳ Follow the thread