NVIDIA has announced the OpenShell Reference Architecture, a comprehensive security framework designed to help enterprises control autonomous AI agents operating across infrastructure and local systems. The architecture divides processing across NVIDIA’s platform, utilizing Vera Rubin systems for AI models and agent software, and BlueField-4 data processing units running Sentry security software. Operating at the hardware level, the system can independently monitor agent behavior, enforce strict security policies, and isolate or stop unauthorized activities within milliseconds. The reference architecture supports cloud environments from AWS, Google Cloud, and Microsoft Azure, alongside local workstations and enterprise integrations like ServiceNow. Designed to address rising concerns over agent breakout risks, the framework aims to give businesses robust oversight while safely scaling automated workflows.

Full report : NVIDIA’s OpenShell Reference Architecture is designed to help businesses control AI agents.