Uber ADR (Agentic AI Detection and Response)

What it is

The open-source release of Uber's ADR includes four of the five described components — Discovery, Sensor, ADR-Bench, and the Detector — but not Prevention (stopping unsafe actions) or the offline ADR Explorer red-teaming engine that hardens detection pre-deployment. So this is an inventory + telemetry + benchmark + detection toolkit, not an enforcement product.

How it works

The default detector, adr, is a two-tier dual-agent design: high-recall triage first, then deeper agentic reasoning only on suspicious sessions. Running the real detector needs both an Anthropic and an OpenAI key. A keyless llamafirewall detector exists for smoke tests. Sensor observability covers 7+ AI coding tools across macOS, Linux, and Windows.

Preconditions / Affected

Relevant if you have employees running agents like Cursor, Claude Code, Codex, GitHub Copilot CLI, or DeepSeek Harness, or customer-facing support agents, and want visibility rather than blocking. Detection requires API keys (spend) unless you use the keyless baseline.

What to do

Setup:

git clone https://github.com/uber/ADR
cd ADR/Detection
uv sync
export ANTHROPIC_API_KEY="..." OPENAI_API_KEY="..."

For a keyless smoke test, run with --detector llamafirewall (see Detection/README.md). The full evaluation workflow — inflate the packed benchmark, run detectors, plot figures — is in docs/REPRODUCIBILITY.md. Component entry points:

• Discovery/README.md — endpoint inventory, probes, fingerprint catalog
• Sensor/README.md — telemetry collection and unified schema
• Detection/README.md — ADR-Bench, detector baselines, MCP infrastructure

Caveats

Apache 2.0, but Detection/benchmark/agentdojo/ is vendored MIT third-party code under its own license. The benchmark ships synthetic fixtures — fake credentials, emulated environments, prompt-injection scenarios — labeled for defensive research only (docs/OPEN_SOURCE_REVIEW.md). No enforcement/prevention is included, so ADR here detects and reports; acting on findings is on you.