NVIDIA Launches OpenShell Architecture for Secure AI Agent Deployment

NVIDIA announced the OpenShell Reference Architecture on Wednesday at GTC Berlin to enable organizations to safely deploy AI agents at scale. The blueprint provides a framework for keeping corporate AI agents within defined boundaries. NVIDIA CEO Jensen Huang stated that the system allows agents to perform useful work while remaining inside clearly defined limits. The architecture calls for NVIDIA Vera Rubin infrastructure to run models such as NVIDIA Nemotron.

The company also introduced Confidential Bare Metal as a new deployment option for its Confidential Computing infrastructure. This expansion supports NVIDIA Grace Blackwell and Rubin systems. Cohere plans to integrate this option into its Cohere Model Vault using trusted execution environments. Google Cloud is extending support for Confidential Bare Metal to A4X Max instances powered by NVIDIA GB300 NVL72 systems. Support for these deployments now extends to CoreWeave, Firmus, GMI Cloud, IREN, Lambda, Microsoft Azure Infrastructure, Nebius, Nscale, and Vultr.

SAP is integrating the OpenShell Reference Architecture with its Joule Studio runtime for safer agent operations. ServiceNow announced general availability of its Arc agent, which has actions governed at runtime and enforced by OpenShell. Sentry inspects agent traffic from a separate hardware domain and can quarantine an agent within milliseconds. NVIDIA lists Sentry's planned availability as early next year. The company noted that several banks, including BBVA and Santander, are perusing the specification. OpenShell collects telemetry data by default to improve the project, though users can opt-out.