Nvidia released an open-source security toolkit designed to prevent AI agents from escaping containment or behaving unpredictably. The tool arrives after a series of high-profile agent safety incidents and is meant as a guardrail layer for enterprise agentic deployments.