Nvidia has expanded its OpenShell security framework to cover full-stack infrastructure, aiming to prevent AI agents from going rogue. The company unveiled the updated reference architecture at its GTC event in Germany, extending OpenShell's policy-based security, network, and privacy guardrails to enterprise workloads, CPUs, DPUs, and edge-based systems like DGX Spark and Station.
Ian Buck, Nvidia’s VP and GM of hyperscale and high-performance computing, emphasized the need for clear runtime boundaries and visibility into agent behavior. The extension includes support for Vera Rubin infrastructure, Vera CPUs, and BlueField DPUs, ensuring security across retail, telecom, and robotics applications.
OpenShell, which debuted earlier this year, initially focused on securing Nemoclaw, Nvidia’s OpenClaw integration. The framework has since been expanded to monitor agents executing tasks across both open and closed models, forming a key part of the Open Agent Safety platform launched after several rogue agent incidents.
The reference architecture is compatible with enterprise operating systems like Linux and Kubernetes, as well as cloud operating systems from major hyperscalers. It also supports models from AI labs including Anthropic, Cohere, and OpenAI, which notably did not participate in Nvidia’s Open Agent Safety group or the Open Secure AI Alliance.