uber

ADR: Detect and Respond to Enterprise AI Agent Threats

ADR is Uber’s Python security system for enterprise AI agents, including coding tools and customer-facing agents. It combines endpoint discovery, agent telemetry, security benchmarks, and threat detection across macOS, Linux, and Windows.

ActivePythonApache-2.0AI AgentsSecurityObservability
Stars
2.1k
Forks
231
Stars, last 30 days
+75
Last commit
1 month ago
Contributors
21

Our take

Activeregular commits and releases

Health score

62/100

How it is scored
Activity
65
Community
66
Issues
58
Pull requests
59

+75 stars in the last 30 days

  • Merges more pull requests than 83% of the projects we track

ADR is an actively developed, Apache-2.0-licensed security toolkit with broad agent telemetry coverage, but its open-source scope stops short of prevention and its default detector needs external model API credentials.

Good fit if

  • You need discovery or observability for AI agents across macOS, Linux, and Windows.
  • Your security team wants to evaluate agent defenses using ADR-Bench or experiment with its detector.
  • You can work with the documented setup and, for the default detector, provide Anthropic and OpenAI API keys.

Look elsewhere if

  • You need an open-source component that blocks unsafe agent actions, which ADR says is not included.
  • You need the ADR Explorer pre-deployment red-teaming engine, which is also excluded from this repository.
  • You need a ready-made detector run without model API credentials; the README identifies LlamaFirewall as a keyless smoke-test option, not the default detector.
All health signals
Last commit2026-10-16 (1 month ago)
Commits, last 90 days100+
Releases, last 12 months1 (latest sensor-v1.0.0, 2026-07-31)
Contributors21 (top contributor: 34% of commits)
Issues closed, last 90 days5+ (typically closed in 3 days)
Pull requests merged, last 90 days74 (typically merged in 1 day)
Project age7 months

Checked on 2026-11-17 with the GitHub API.

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

2,140

Forks

231

Change since Nov 17, 2026: 0 stars·0 forks

Overview

ADR (Agentic AI Detection and Response) is a Python enterprise security system for AI agents, including employee-facing tools such as Cursor, Claude Code, and Codex, and customer-facing support agents. Deployed in production at Uber, it helps security teams find unsanctioned tools, understand agent activity, evaluate defenses, and detect risky behavior across macOS, Linux, and Windows.

The open-source repository includes ADR Discovery, Sensor, Benchmark, and Detector. It is aimed at organizations that need visibility and security evaluation for agents in use, rather than a general-purpose agent framework. Prevention and the offline ADR Explorer engine are not included in this release.

Key Features

  • Inventories AI applications, CLI agents, IDE extensions, local model runtimes, and MCP servers on endpoints.
  • Collects and normalizes agent telemetry, including intent, tool use, and execution traces.
  • Supports telemetry from coding tools and agent applications including Claude Code, Cursor, Codex, GitHub Copilot CLI, DeepSeek Harness, and Gemini CLI.
  • Provides ADR-Bench for evaluating agent security with tasks, MCP servers, and attack-technique coverage.
  • Uses a two-tier detector, with high-recall triage followed by deeper agentic reasoning for suspicious sessions.
  • Includes a reproducibility workflow for running benchmark detectors and generating paper figures.

Use Cases

  • Security teams can inventory AI tools and MCP servers installed across employee endpoints, then review unfamiliar surfaces.
  • Incident responders can inspect agent traces to understand which tools an agent used and what it did.
  • AI platform teams can benchmark agent defenses against realistic enterprise-style tasks before deployment.
  • Teams operating customer support agents can explore detection approaches for risky agent behavior alongside employee-facing use cases.

What you need

Detected in the repository

  • Python (main language on GitHub)
  • Automated checks on GitHub Actions

License in plain words

Apache-2.0permissive

  • Commercial use: yes
  • Modify and redistribute: yes
  • You must keep: the license, the NOTICE file and a note of your changes
  • Share your changes: no
  • Includes an explicit patent grant from the contributors:

A summary, not legal advice: the LICENSE file is what applies.

Getting Started

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

See the Detection README and reproducibility guide for detector options and evaluation steps.

Alternatives

  • iFixAi: iFixAi audits agent behavior in simulated business workflows, while ADR focuses on endpoint discovery, telemetry, benchmarks, and threat detection across deployed agents.
ProjectLanguageLicenseStarsStatus
ADRPythonApache-2.02.1kActive
iFixAiPythonApache-2.045.2kActive

Considerations

The project is active, with regular commits and merges, and contributions from multiple people rather than a project dominated by one contributor. Its recent release history is relatively limited, and the measured issue data indicates issues are generally closed quickly, though only a small number were closed in the measured period. Treat it as a developing security system and validate its behavior against your own agents and policies.

The default detector setup calls for both Anthropic and OpenAI API keys; the README offers LlamaFirewall for keyless smoke tests. ADR Prevention and ADR Explorer are not part of the open-source release. Benchmark fixtures are synthetic and intended for defensive security research. The project is licensed under Apache-2.0, while vendored AgentDojo code has its own MIT license.

Found this useful?

Share it with someone who would like ADR.

Comparisons