ADR
Secures enterprise AI agents through discovery, observability, threat detection, and security benchmarking.
Monitor endpoint AI tools, evaluate agent security risks, and detect attacks using a dual-agent architecture.
ADR is an enterprise security framework developed by Uber to detect and monitor AI agents running across employee endpoints and customer-facing support tools. You should use it if you need visibility into shadow AI software, MCP servers, and agentic prompt injection attacks. It provides production-tested telemetry collection and threat benchmarking for infrastructure security teams.
What it is
ADR inventories installed AI applications, captures agent execution traces across tools like Cursor and Claude Code, and runs threat benchmarks across 17 attack techniques. Its dual-agent detection engine combines initial high-recall filtering with deep reasoning on suspicious sessions. Note that automated action prevention and offline red-teaming components are excluded from this open-source release.
Who it's for
This system suits enterprise cybersecurity teams, IT infrastructure administrators, and security researchers auditing AI agent behavior on corporate endpoints. It does not suit individual developers seeking a lightweight firewall or non-technical teams unable to manage Python environments and LLM API integrations.
Key features
Endpoint AI Discovery
Inventories local AI applications, CLI agents, IDE extensions, runtimes, and Model Context Protocol servers to flag unknown surfaces.
Cross-Platform Agent Observability
Captures normalized telemetry, tool execution, and user intent across major AI development tools on macOS, Linux, and Windows.
Security Attack Benchmarking
Evaluates defenses against 17 agent attack techniques using over 300 synthetic tasks and 134 MCP server environments.
Dual-Agent Threat Detection
Triages potential threats with lightweight filters before routing suspicious sessions to deeper LLM-based reasoning engines.
Up and running in 3 steps
Our honest verdict
What's great
- Proven at production scale at Uber and accepted at MLSys 2026.
- Captures normalized telemetry across major AI tools including Cursor, Claude Code, Codex, and Gemini CLI.
- Includes over 300 benchmark tasks and 134 MCP server environments for defensive testing.
- Supports keyless offline testing using Llama Firewall baselines.
Where it falls short
- Action prevention capabilities are not included in the open-source codebase.
- Running the primary dual-agent detector requires active Anthropic and OpenAI API keys, incurring usage costs.
- Requires installing endpoint sensors across your fleet and managing Python dependencies via uv.
- The offline ADR Explorer engine for automated pre-deployment red teaming is excluded.
Pricing & license
ADR is free and open-source under the Apache-2.0 license. The primary cost comes from running the underlying compute and paying for Anthropic and OpenAI API requests consumed during detection runs.
Open remote AI-training roles
873 open remote AI-training roles, updated November 15, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Does this open-source release automatically block malicious AI agent actions?
No, the prevention component that stops unsafe actions before execution is omitted from the open-source release.
What operating systems does the ADR Sensor support?
The sensor collects agent session journals across macOS, Linux, and Windows platforms.
Can I test the detection engine without commercial LLM API keys?
Yes, keyless smoke tests can be executed using the llamafirewall detector flag inside the Detection directory.
