NVIDIA unveils OpenShell Reference Architecture, an architecture for deploying AI agents safely
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
NVIDIA has brought its data center hardware and its security software together in a reference design built so companies can put AI agents to work under control. Sentry tracks each agent's activity from BlueField-4 chips and isolates it in milliseconds if it oversteps its permissions. Anthropic, OpenAI and Microsoft are collaborating on the project, and BBVA and Santander are studying it.
At its GTC Berlin conference on Wednesday, October 21, NVIDIA presented OpenShell Reference Architecture: a reference hardware and software design meant to help companies create, isolate, watch over and control AI agents, both in their own infrastructure and on local machines.
The architecture starts from OpenShell, NVIDIA's open source software that sets the limits within which an agent may act. The company announced it on September 28, together with NVIDIA Sentry, as part of Open Agent Safety Platform, and now combines them with its data center systems: Vera Rubin infrastructure runs the models, Vera processors run the agents and their tools, and BlueField-4 data processing units run Sentry, which keeps a constant, separate watch on each agent's activity. When an agent tries to step past its limits, Sentry puts it in quarantine and halts it within milliseconds.
Outside the data center, NVIDIA extends it to its RTX Spark, DGX Spark and DGX Station machines, with examples in sectors such as healthcare, retail, robotics, manufacturing, telecommunications and software development. It runs on Linux and Kubernetes from Canonical, SUSE and Red Hat, on Windows with Microsoft Execution Containers, and in the AWS, Google Cloud and Microsoft Azure clouds.
Microsoft, Anthropic, AWS, Meta, Google Cloud and OpenAI are among the companies working with this architecture, and European banks such as Santander and BBVA are exploring it with NVIDIA to strengthen the security and control of their agents. SAP is building it into Joule Studio, part of its SAP Business AI platform, and on the same day ServiceNow announced ServiceNow Arc, an agent whose limits are enforced by OpenShell.
"AI capability and safety must advance together," said Jensen Huang, NVIDIA's founder and CEO.
According to NVIDIA, the architecture is now available. The release gives no prices.
Why it matters · analysis and opinion
Until now, putting limits on an agent depended mostly on the application that runs it, and NVIDIA itself acknowledged in September that recent incidents came from agents getting around those controls. Moving the monitoring to a separate chip, out of the agent's reach, changes the approach: security stops being a layer added at the end. Having banks and major labs on board suggests that what holds back agents in sensitive tasks is no longer so much what they can do as the confidence that they will not overstep. The release gives no prices, and the most striking piece, Sentry, is tied to NVIDIA's own chips. Anyone testing agents can start with what does not require buying hardware: writing down what each one is allowed to touch and reviewing its logs.
Official source: NVIDIA · Written with the help of AI: how we make the news
