NVIDIA launches OpenShell reference architecture to provide security protection for enterprise-grade AI agents.

October 21: Nvidia has launched the OpenShell reference architecture, integrating Vera Rubin infrastructure, Vera central processing units (CPUs), BlueField-4 data processing units (DPUs), and OpenShell secure runtime software to help enterprises build, run, and manage AI agents at scale, while monitoring and enforcing security constraints on their behavior.

The architecture continuously monitors agent behavior through the NVIDIA Sentry security framework running on BlueField-4 and enforces security policies at the hardware level. When an agent attempts to breach software-defined boundaries, the system can isolate and halt it within milliseconds. Nvidia claims that the Vera Rubin NVL72 system can achieve up to 30 times the throughput per megawatt under agent workloads compared with previous systems, while NVIDIA Sentry can increase agent hosting capacity by up to 20%.

The architecture supports enterprise environments such as Linux, Kubernetes, and Microsoft Windows, and can scale to edge devices. Companies including Microsoft, Google Cloud, Dell Technologies, Cisco, OpenAI, and Anthropic are participating in related collaborations.