uber/ADR

ADR addresses the lack of visibility and security controls for AI agents in enterprise environments. It solves the problem of detecting and preventing malicious or unsafe agent behavior, which traditional security tools are not designed to handle, by providing a unified telemetry schema, a realistic benchmark for testing defenses, and an efficient two-tier detection architecture.

Developers can build comprehensive security monitoring and response systems for AI agents, including real-time observability of agent actions, automated threat detection, and integration with existing enterprise security workflows. The framework supports a wide range of agents (e.g., Cursor, Claude Code, Codex) and can be extended to custom agents. The benchmark enables continuous evaluation and hardening of agent security postures, and the detection system can be deployed to flag risky sessions in production.