Nvidia wraps OpenShell around full-stack silicon to stop AI agents going rogue

Nvidia unveiled an OpenShell reference architecture at GTC in Germany that extends its open-source agent runtime from system-layer guardrails to full-stack infrastructure, covering Vera Rubin systems for enterprise workloads, Vera CPUs running agent harnesses, BlueField data processing units for security frameworks, and edge systems including DGX Spark and Station. The architecture runs on enterprise operating systems and container orchestration layers including Linux and Kubernetes distributions such as SUSE and Red Hat, plus major hyperscaler cloud operating systems, and supports models from Anthropic, Cohere and OpenAI. Nvidia VP and GM of hyperscale and high-performance computing Ian Buck said long-running agents "combine tools, credentials, and network access in ways that traditional security controls were not designed to govern," while CEO Jensen Huang framed OpenShell as giving agents "the freedom to do useful work within clearly defined boundaries.

BERLIN – Nvidia looks to continue its quest to shore up security for agentic AI, unveiling a reference architecture based on its open-source runtime software, OpenShell. Where initial OpenShell enforces policy-based security, network, and privacy guardrails for agents at the system layer, the concept unveiled at GTC in Germany extends that to full-stack infrastructure. The extension sees OpenShell now covering Vera Rubin infrastructure for running enterprise workloads, Vera CPUs operating agent harnesses, and BlueField data processing units for running security software frameworks. It also encompasses edge-based systems like DGX Spark and Station for agents operating across retail, telecoms, and robotics. “At Nvidia, we take our role as stewards of AI very seriously, Ian Buck, Nvidia’s VP and GM of hyperscale and high-performance computing, said in a press briefing. “AI capabilities and safety must continue to advance together. Long-running agents combine tools, credentials, and network access in ways that traditional security controls were not designed to govern. So, protecting these workloads requires really clear runtime boundaries, visibility into the agent behavior, and independent enforcement across the infrastructure stack that they run on.” OpenShell debuted at Nvidia’s flagship GTC event in the Spring, providing security guardrails for Nemoclaw, Nvidia’s OpenClaw integration offering. The concept was later expanded to provide boundaries for agents executing tasks across both open and closed models, while also forming a central part of the company’s community-based Open Agent Safety platform https://www.sdxcentral.com/news/nvidia-ropes-in-100-strong-posse-to-leash-rogue-ai-agents-after-openais-walkabout/ launched in the wake of several rogue agent incidents from the likes of OpenAI. The dedicated OpenShell Reference Architecture can run on enterprise operating systems and container orchestration layers, including Linux and Kubernetes infrastructure such as SUSE and Red Hat, as well as cloud operating systems from major hyperscalers. It also covers models from AI labs and builders including Anthropic, Cohere, and most notably OpenAI – which was a notable absentee from both Nvidia’s Open Agent Safety https://www.sdxcentral.com/news/nvidia-ropes-in-100-strong-posse-to-leash-rogue-ai-agents-after-openais-walkabout/ group and the Open Secure AI Alliance https://www.sdxcentral.com/news/broadcom-joins-nvidia-ai-security-alliance-we-havent-forgotten-open-source/ . “AI capability and safety must advance together,” Jensen Huang, Nvidia’s founder and CEO, said. “As enterprises prepare to deploy billions of AI agents, safety and control must be built into the computing architecture. Nvidia created OpenShell to give agents the freedom to do useful work within clearly defined boundaries – contained, monitored, and under control.”