NVIDIA unveils OpenShell architecture to secure AI agents

At the GTC Berlin conference, NVIDIA announced the OpenShell Reference Architecture, a full-stack solution designed to build, isolate, observe, and govern autonomous AI agents within enterprise and edge environments. The architecture integrates several key hardware and software components: NVIDIA Vera Rubin infrastructure for model inference, NVIDIA Vera CPUs for executing agent harnesses and tool calls, and NVIDIA BlueField-4 DPUs running the NVIDIA Sentry security framework.

NVIDIA Sentry provides in-silicon policy enforcement and out-of-band monitoring, allowing the system to quarantine or block an agent within milliseconds if it attempts to exceed its defined software boundaries. This hardware-level security is intended to protect against compromised host environments.

The architecture is being adopted by a wide range of industry leaders, including Anthropic, AWS, Cisco, CrowdStrike, Google Cloud, IBM, Meta, Microsoft, OpenAI, SAP, and ServiceNow. Additionally, companies like BBVA and Santander are collaborating with NVIDIA to explore or implement these technologies for specialized banking tasks, such as fraud prevention and customer service automation. SUSE has also announced integration of the architecture into its AI Factory to provide sovereign control over agentic AI workflows.