Securing the agentic future. Open-source, research-driven security for AI agents - runtime monitoring, detection, and control for systems that don't behave like traditional software.
Adrian is our open-source runtime AI agent security tool. It analyses an agent's tool calls and reasoning traces, then catches malicious tool use, prompt injection, and out-of-remit behaviour in real time - and can intervene in-flight (audit or block mode). AARM-aligned, Apache-2.0, fully self-hostable offline. Two-line LangChain/LangGraph install.
We publish on agent security, red-teaming, and the agentic threat landscape: