ADR

ADR (Agentic Detection and Response) is the research artifact for a paper, "ADR: An Agentic Detection System for Enterprise Agentic AI Security" (MLSys 2026, arXiv 2605.17380). The README describes it as an open-source telemetry sensor and a dual-agent benchmark and detection framework for securing AI agents that use the Model Context Protocol. This is defensive security work.

The repository has two parts. `Detection/` holds the ADR Detector and ADR-Bench: a dual-agent detector, 133 MCP servers, 303 benchmark tasks, baselines and figure scripts. `Sensor/` collects and normalizes agent telemetry from Claude Code, Cursor, Codex and others into a unified schema. A reproducibility guide walks through inflating the packed benchmark, running detectors and plotting the paper's figures. The README says the offline ADR Explorer red-teaming engine described in the paper is not included.

To try it: clone the repo, `cd ADR/Detection`, `uv sync`, and set `ANTHROPIC_API_KEY` and `OPENAI_API_KEY`; the default detector is `adr`, and a keyless smoke test is possible with `--detector llamafirewall`. A data notice says the benchmark uses synthetic fixtures (fake credentials, emulated environments, prompt-injection scenarios) for defensive research. The license is Apache-2.0, with vendored AgentDojo code under MIT, and the last push was 2026-10-12. This review is based on the README and repository data; we have not run it ourselves.