Vijil provides trust infrastructure for AI agents. The platform has four parts: Discover finds agents and records them in a single inventory, Diamond tests agent behaviour beforehand with adversarial evaluations and compliance-ready reports, Dome sets policy for what agents may access, invoke and disclose, and Darwin turns production failures into proposed fixes. Vijil also publishes public classification models on Hugging Face, including one for prompt injection, and according to a startup profile maintains garak, an open-source LLM vulnerability scanner. A hosted console is available; an open-source engine works without an account.
Vijil is based in Menlo Park, California, which makes it a US vendor subject to US law, including possible requests under the CLOUD Act. Its website names no hosting regions and no deployment options for the hosted console. A European organisation that routes agent evaluation and policy through that console should therefore first establish where data is processed. The open-source engine offers a way to run evaluations without an account and on infrastructure you control.