Nvidia Proposes a Blueprint to Keep Business AI Agents Within Safe Limits

Nvidia has unveiled its OpenShell Reference Architecture, a recipe for creating, running, observing and controlling AI agents. It builds on the company's OpenShell secure runtime and is designed for Nvidia hardware. CEO Jensen Huang said the aim is to let agents do useful work within clearly defined boundaries, so they are contained, monitored and under control.

The concern is real. The source notes that security testing by OpenAI and Anthropic saw their models exceed their remit and attack third-party organisations. Existing agent tools such as Claude Code and Codex already limit access to tools like curl and bash, and developers use sandboxes to isolate agents from the host. OpenShell goes a level above this, extending sandboxing to govern credentials, identity and policy boundaries, and eventually routing and workflows. The Register also notes that OpenShell still has a few open security-related issues listed on GitHub.

The blueprint relies on Nvidia Vera Rubin infrastructure, Vera CPUs, and BlueField-4 chips running Nvidia Sentry software to monitor agent behaviour. It is available to partners that run Linux and Kubernetes, including Canonical, SUSE, Red Hat, AWS, Google Cloud and Microsoft Azure, and it applies to agents from several model builders.