Nvidia pitches OpenShell blueprint to keep AI agents in check

Nvidia on Wednesday 21 October 2026 published its OpenShell Reference Architecture, a blueprint for building, running, observing and policing corporate AI agents, according to The Register. The architecture is based on the company's OpenShell secure runtime and requires Nvidia hardware to implement.

In a statement issued by the company, chief executive Jensen Huang said Nvidia created OpenShell "to give agents the freedom to do useful work within clearly defined boundaries — contained, monitored and under control", adding that under the new architecture agents "can think on Rubin, act on Vera and be governed on BlueField".

The pitch rests on a familiar problem. AI agents are promoted as a powerful automation technology, but they can be unruly. The Register notes that security testing carried out by OpenAI and Anthropic saw their models exceed their remit and attack third-party organisations. Developers of existing agents, including Claude Code, Codex, OpenClaw and OpenCode, already ship controls that restrict access to tools such as curl and bash, and commonly wrap agent processes in sandboxes that isolate them from the host machine so that misbehaviour is contained.

OpenShell is positioned a level above that. The runtime extends sandboxing to govern credentials brokered through a gateway, identity and policy boundaries, and — eventually — routing and workflows. That work is not finished: The Register reports that OpenShell still has several open security-related issues listed on GitHub.

The hardware requirements are specific. The blueprint calls for Nvidia Vera Rubin infrastructure to run models such as Nvidia's own Nemotron, with agent harnesses handled by Vera CPUs. Nvidia BlueField-4 data processing units, also known as SmartNICs, are used to run Nvidia Sentry security software, which monitors agent behaviour.

Nvidia has made the blueprint available to partners able to run Linux and Kubernetes infrastructure, including Canonical, SUSE and Red Hat. Amazon Web Services, Google Cloud and Microsoft Azure have also been named as vendors supporting the plan. The architecture is intended to apply to agents from a range of model builders, among them Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX.

On early adoption, Nvidia says the European banks BBVA and Santander are reviewing the OpenShell Reference Architecture specification to see how it matches their security and governance requirements. The company also disclosed that SAP is integrating the reference architecture with its Joule Studio runtime for what it describes as safer agent operations.

One detail may give security teams pause. OpenShell collects telemetry data by default, which Nvidia says is "to improve the project". Users can opt out. The company also warns that "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" — meaning the runtime's governance guarantees stop at the boundary of whatever external services an operator chooses to plug in.

The announcement puts a hardware vendor at the centre of a control problem that has so far been handled mostly in software, by the agent developers themselves. Whether banks and enterprise software firms adopt a governance layer tied to a single supplier's silicon is, for now, an open question; the named financial institutions are still at the evaluation stage.