Nvidia has unveiled the OpenShell reference architecture for controlling enterprise AI agents, The Register reported on Tuesday.

OpenShell is an AI agent security runtime environment developed by Nvidia. The OpenShell reference architecture describes how to build and operate AI agents with OpenShell on Nvidia hardware. It also includes ways to monitor what agents do and prevent them from going beyond set boundaries.

Nvidia CEO Jensen Huang (젠슨 황) said, "OpenShell was built to make agents do useful work within clear boundaries." He said, "Agents think on Rubin, act on Vera and are controlled on BlueField."

AI agents can also escape controls. In security tests conducted by OpenAI and Anthropic, there were cases in which models exceeded their authority and attacked external organisations.

OpenShell goes beyond a sandbox by extending management to credential, identity and policy boundaries brokered by a gateway. It is also expected to support management of routing and workflows in the future.

Using OpenShell requires Nvidia hardware.

Canonical, SUSE and Red Hat, as well as Amazon Web Services, Google Cloud and Microsoft Azure, joined as partners for the OpenShell reference architecture.

Nvidia said European banks BBVA and Santander are considering adopting the OpenShell reference architecture. SAP is integrating the OpenShell reference architecture into its Joule Studio runtime.

OpenShell collects usage log data by default. Nvidia explained that this is to improve the project. Users can refuse collection in settings.