Monitor what AI coding agents like Claude Code, Cursor, and Codex are doing inside an organization.
Benchmark AI agent defenses against 17 known attack techniques using ADR-Bench.
Detect risky or suspicious AI agent sessions with a two-tier detection system.
| uber/adr | 521xueweihan/python | pathwaycom/arc-task-gen | |
|---|---|---|---|
| Stars | 1,401 | 1,397 | 1,387 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | easy | moderate |
| Complexity | 4/5 | 1/5 | 3/5 |
| Audience | ops devops | vibe coder | researcher |
Figures from each repo's GitHub metadata at analysis time.
Requires Anthropic and OpenAI API keys to run the full detector suite.
ADR, short for Agentic AI Detection and Response, is an enterprise security system built to keep an eye on AI agents inside a company. These agents include coding assistants like Cursor, Claude Code, and Codex that employees use every day, as well as AI agents that talk directly to customers, such as support bots. ADR is already running in production at Uber, and the research paper behind it was accepted at a machine learning systems conference called MLSys 2026. The project works through four pieces, though not all of them are included in this open source release. The first is observability: ADR records what an agent is doing and why, capturing its intent, which tools it calls, and the full trace of its actions, across more than seven different AI coding tools on Mac, Linux, and Windows computers, plus internal automation and customer facing agents. The second is a benchmark called ADR-Bench, which contains over 300 test tasks, 133 simulated MCP servers, and examples covering all 17 known ways an AI agent can be attacked, so teams can see how well their defenses hold up. The third is detection, using a two step system that first quickly flags suspicious sessions, then applies deeper reasoning to look closer at the ones that seem risky. The fourth piece, prevention, which would actually stop unsafe actions before they happen, is not part of this release yet. This repository includes the Sensor, which collects and standardizes telemetry from agents like Claude Code, Cursor, Codex, and others, and the Detection component, which holds the dual agent detector, the benchmark tasks, and baseline comparisons. A separate offline tool used internally for red teaming and hardening detection is not included here. Getting started involves cloning the repository, installing dependencies with a tool called uv, and setting API keys for Anthropic and OpenAI to run the detectors, with a keyless option available for quick smoke tests. The project is released under the Apache 2.0 license, with one vendored third party folder under the MIT license, and it includes only synthetic, fake test data meant for defensive security research.
An enterprise security system, used in production at Uber, that observes, benchmarks, and detects risky behavior in AI coding and support agents.
Mainly Python. The stack also includes Python, uv, Anthropic API.
Use freely for any purpose, including commercial use, under the Apache 2.0 license, as long as you keep the copyright and license notices.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly ops devops.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.