Built on the OpenShell secure runtime, Nvidia’s reference architecture manages agent identities, credentials and policies, with Canonical, Red Hat, SUSE and the three major cloud providers as partners.
Nvidia unveiled Wednesday its OpenShell Reference Architecture, a kind of recipe to create, operate, observe and enforce rules for AI agents deployed in companies. It is based on OpenShell, the group's secure runtime, which goes beyond conventional sandboxes: it also manages identifiers transmitted via a gateway, identities and access policy limits, and should eventually cover routing and workflows. The context is known: the security tests conducted by OpenAI and Anthropic showed agents leaving their perimeters to the point of attacking third-party organizations.
Security through Nvidia equipment
The scheme assumes an end-to-end Nvidia infrastructure: models (such as Nemotron) run on Vera Rubin, agent harnesses on Vera processors, and BlueField-4 DPUs host Nvidia Sentry surveillance software. Canonical, SUSE, Red Hat, AWS, Google Cloud and Microsoft Azure are among the partners, and the agents of many model publishers are involved. The BBVA and Santander banks are studying the specification, and SAP integrates it into its runtime Joule Studio.
The Register notes two flats: some security issues remain open on OpenShell's GitHub repository, and the tool defaults to send telemetry data to Nvidia, a collection that can be disabled.
Read the full article on the original source
Go to the source →