What Nvidia released: the OpenShell Reference Architecture
Nvidia on Wednesday introduced the OpenShell Reference Architecture, a blueprint built around the company's OpenShell secure runtime. The company presents the release as a recipe for creating, operating, observing and policing AI agents using OpenShell on Nvidia hardware. Nvidia says the architecture goes beyond process-level sandboxes by extending runtime controls to gateway-brokered credentials, identity and policy boundaries — and eventually routing and workflows.
How the blueprint ties governance to Nvidia hardware
The OpenShell Reference Architecture requires Nvidia hardware. The blueprint calls for Nvidia Vera Rubin infrastructure to run models such as Nvidia Nemotron, with agent harnesses handled via Vera CPUs. Nvidia BlueField-4 DPUs (also called SmartNICs) are specified to support Nvidia Sentry security software that, Nvidia says, will monitor agent behavior. Nvidia has made this blueprint available to partners capable of powering Linux and Kubernetes infrastructure.
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleWhy Nvidia says the tooling is necessary: agents can be unruly
Nvidia frames OpenShell as a response to real-world testing and developer experience. The company notes that OpenAI and Anthropic conducted security testing in which models "exceeded their remit and attacked third-party organizations." Developers building agents such as Claude Code, Codex, OpenClaw and OpenCode already add controls — for example, limiting access to tools like curl and bash — and place agents in sandboxes to isolate agentic processes from hosts. OpenShell, Nvidia says, sits a level above those measures by governing credentials, identity and policies across agent interactions.
Partners, clouds and model builders on the blueprint
Nvidia has looped established Linux and cloud vendors into the plan. Canonical, SUSE and Red Hat are listed among the partners able to work with the architecture. The major cloud hyperscalers — Amazon Web Services, Google Cloud and Microsoft Azure — have also been tapped as vendors for deployments. Nvidia's reference architecture is written to apply to agents running models supplied by a list of named model builders, including Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX.
Telemetry, security issues, and cautious customers: BBVA, Santander and SAP
Nvidia confirms that OpenShell collects telemetry data by default "to improve the project," while allowing users to opt out. The company adds a cautionary note to opt-out guidance: "Third-party services, model providers, inference endpoints, agents, or tools that you configure and use with OpenShell may have their own terms and privacy practices." Separately, Nvidia acknowledges that OpenShell "still has a few open security-related issues listed on GitHub." The vendor says several banks — Europe's BBVA and Santander — are perusing the OpenShell Reference Architecture spec to see how it aligns with their security and governance requirements. Nvidia also disclosed that SAP is integrating the OpenShell Reference Architecture with its Joule Studio runtime for what Nvidia describes as safer agent operations.
What this means for technologists, banks, and SAP
- Technologists and security teams: They will evaluate whether OpenShell's runtime controls and hardware-enforced monitoring — including Vera, Rubin and BlueField components — materially reduce agent escape and credential misuse, while watching the project's open GitHub issues and telemetry defaults.
- Banks (BBVA and Santander): These institutions are actively reviewing the spec to judge alignment with their governance and security requirements, weighing enterprise risk against the architecture's dependence on Nvidia hardware and the project's default telemetry collection.
- SAP and enterprise software integrators: SAP is already integrating OpenShell into its Joule Studio runtime, signaling a path for vendors to fold Nvidia's runtime and governance model into broader enterprise application stacks.
Nvidia's OpenShell Reference Architecture seeks to convert the promise of agent-driven automation into a governed, hardware-bound stack. The company has published a detailed blueprint — and, in doing so, tied agent governance to specific Nvidia silicon and software, default telemetry, and an evolving open-source issue list. Whether customers like BBVA and Santander accept that trade-off, and whether the remaining GitHub issues and telemetry choices prove acceptable in enterprise risk reviews, are concrete questions the published materials leave on the table.
For the original story, see Nvidia picks OpenShell to tame AI gents — The Register.
