Adrian
Adrian is an open-source runtime security monitor for AI agents that watches both an agent's actions and its reasoning traces to catch prompt injection, tool poisoning, data exfiltration and out-of-remit behavior as it happens, and can intervene before the action completes.
Adrian analyzes both an agent's activity logs, tool calls, actions and outputs, and its reasoning traces, so it can catch not just what an agent did but why it decided to do it, under what context, and what it plans next. It is aligned to a stated Agent Activity and Reasoning Model framework, and its authors cite outside research finding that combining behavior and reasoning analysis catches substantially more attacks than watching behavior alone.
It catches prompt injection and jailbreaks, both direct and indirect, tool poisoning and unsafe or off-policy tool calls, data exfiltration and secret or credential leakage, and privilege escalation or actions outside an agent's intended remit, and can optionally intervene in-flight rather than only reporting after the fact. Software development kits are available for Python, with LangChain and Anthropic integrations, and for TypeScript, and a native Claude Code plugin classifies every tool call in real time, blocking or holding risky actions for approval directly in the terminal.
It is written in Python and released free under an open-source license.