What is adrian?

Adrian is an open-source security monitoring and control system for AI agents, software programs that can take actions autonomously, like browsing the web, calling APIs, or running code. The problem it solves is that AI agents can be manipulated through "prompt injection" (tricking the agent via malicious content it reads), can start behaving in ways that violate their intended purpose, or can make dangerous tool calls. Adrian watches the agent in real time and can block suspicious actions before they happen. What makes Adrian different from simpler monitoring tools is that it analyses not just what the agent does, the tool calls and outputs, but also why it is doing it by examining the agent's reasoning trace. According to the readme, combining behaviour and reasoning analysis improves detection accuracy compared to looking at actions alone. You integrate Adrian into a LangChain or LangGraph agent (two popular Python frameworks for building AI agents) with two lines of Python code. Events are sent to a dashboard where you can review them, classified by severity. You can run in audit-only mode to observe without interfering, or in block mode to halt suspicious actions. The stack can be used through a managed cloud dashboard or self-hosted entirely on your own server using Docker. The self-hosted version runs a Go backend, a web dashboard, and a local AI model for classification. It requires an NVIDIA GPU for the bundled classifier to run at reasonable speed. The project is written in Python for the SDK and is licensed under Apache 2.0.