Summary

Adrian is an open-source runtime security engine for AI agents, detecting and optionally controlling malicious tool use, prompt injection, and policy drift in real-time. It enables confident deployment of agentic AI by providing security monitoring before agents execute actions.

How It Works

Adrian uniquely analyzes both agent activity logs and reasoning traces, combining behavioral and reasoning analysis to improve detection accuracy (~35%) and catch nuanced attacks. It uses "world models" for human-like risk assessment, correlating behaviors and evaluating actions against the agent's intended purpose, moving beyond simple pattern matching.

Quick Start & Requirements

• Managed: Sign up at app.adrian.secureagentics.ai for an API key.
• SDK Install: pip install adrian-sdk and pip install langgraph langchain-openai (or provider). Verified dependencies (May 2026). Integration adds adrian.init()/adrian.shutdown() around agent code.
• Self-hosting: Requires Docker, Docker Compose v2, and an NVIDIA GPU with CUDA/NVIDIA Container Toolkit for the bundled classifier (~10 GB disk). Setup involves cloning, docker compose --profile setup run --rm setup bootstrap (interactive model download), and docker compose --profile llm up -d. Dashboard at http://localhost:3000.
• Docs: Documentation, Quickstart.

Highlighted Details

• Combines behavior and reasoning analysis for ~35% higher detection accuracy and 4x better capture of nuanced attacks (per OpenAI/DeepMind research).
• Employs "world models" for human-like risk assessment, not just pattern matching.
• Provides real-time, in-flight intervention.
• Available free, forever.

Maintenance & Community

Community support via Discord. Contributions welcome per CONTRIBUTING.md.

Licensing & Compatibility

Apache 2.0 license, suitable for commercial use.

Limitations & Caveats

Self-hosting the classifier requires specific NVIDIA GPU hardware and CUDA setup; CPU-only is significantly slower. Verified dependency versions are from May 2026.