These days, an AI agent is not content with just fielding questions; it is taking action in enterprise systems, over networks and on business tools and files. With that expansion of ability comes the risk of an agent overstepping its authority.
To meet that head on, NVIDIA unveiled the OpenShell Reference Architecture at GTC Berlin. It is a full-stack infrastructure platform for the building, monitoring and governance of AI agents in both edge and enterprise settings, complete with hardware controls. For those in business making use of agentic AI, the idea is to have agents do their complex work without having to rely on the model for security and governance.
How NVIDIA’s Three-Layer Architecture Secures AI Agents
“Think on Rubin, act on Vera, and be governed on BlueField,” is how NVIDIA puts its architecture.
The NVIDIA Vera Rubin infrastructure is the part that does the inference on models, open or closed, made with Nemotron. In terms of agentic workloads, the company says the Vera Rubin NVL72 offers 30 times the throughput per megawatt. The Vera CPUs are put to work on agent frameworks and tools, where they are said to outperform x86 by a factor of two while being more energy efficient.
Then there is the BlueField-4 DPU running NVIDIA Sentry. Based on DOCA, it enforces policy at the hardware level and provides out-of-band monitoring. Should an agent go beyond what is allowed in software, Sentry will quarantine and halt it in a matter of milliseconds. NVIDIA also has figures showing Sentry can put up to 20% more capacity in an agent.
How OpenShell Keeps AI Agents Within Defined Limits
As an open-source secure runtime, OpenShell does not trust the agent workloads. It dictates which network services, tools, files and systems are off limits. Four pieces make this happen: the Gateway is in charge of sandbox policies and lifecycles; the Supervisor mediates credentials and looks at outbound requests; the Sandbox puts a wall around workloads to limit process and filesystem access; and the Policy Prover makes sure permissions do not exceed the ceiling set by the organisation.
It will run on Linux and Kubernetes from the likes of Canonical, SUSE and Red Hat, as well as Microsoft Windows via Execution Containers and in the cloud on AWS, Google Cloud and Azure.
NVIDIA Builds a Broad AI and Cybersecurity Partner Ecosystem
The ecosystem is broad. On the model side one finds Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence. There are cloud providers like AWS, Oracle and CoreWeave, system makers such as Dell Technologies, HPE, Lenovo and Supermicro, and cybersecurity firms including CrowdStrike, Cisco, Palo Alto Networks, Zscaler, Okta and Fortinet.
SAP is putting OpenShell into Joule Studio. ServiceNow has made available its Arc personal AI agent, overseen by its Super Intelligence Control Tower and enforced by OpenShell. If need be, ServiceNow’s AI Passport layer can pull an agent’s access from software and hardware in seconds.
From Data Centres to Real-World AI Applications
With RTX Spark, DGX Spark and DGX Station systems the architecture is aimed at the likes of retail, healthcare, manufacturing, robotics and telecoms. BBVA and Santander are looking into the framework, and IBM, SAP, Siemens and others are constructing agent platforms on it.
Jensen Huang, NVIDIA’s CEO, would have it that safety and AI capability must progress in tandem. The OpenShell method is to let the agent have space to operate but see to it that its actions are properly contained and governed. That is the control enterprises want from these ever more able systems.
