Nvidia's Scale-In Architecture Fortifies AI Agent Security in Enterprise Data Centers

Nvidia Corp. has unveiled Scale-In, an addition to its AI networking infrastructure aimed at securing, managing and operating agentic AI factories. The platform delivers secure access, high-performance data movement, tenant isolation, streamlined operations and more consistent performance as AI agents take on larger responsibilities within data center environments.

The shift to agentic AI introduces distinct infrastructure challenges. "In agentic AI, it's a completely different use case because now we're dealing with many agents that need to run on the infrastructure," explained Gilad Shainer, senior vice president of networking at Nvidia. "You need to increase security because you need to make sure that those agents are contained in their sandbox and are not doing the wrong things. You need to build an infrastructure that needs to ensure security all around…what happens in the memory, what happens on storage. This is why we needed to bring an infrastructure which actually enabled the security across everything. And that's what Scale-In is."

Since introducing Scale-In in August, Nvidia has broadened its agent security approach. The company launched its Open Agent Safety Platform in September, merging OpenShell with Nvidia Sentry to provide out-of-band monitoring and hardware-based security enforcement. An October release brought the OpenShell Reference Architecture, which incorporates Vera Rubin infrastructure, Vera CPUs and BlueField-4 to enable secure agent execution across enterprise deployments.

In-silicon security for AI networking

Scale-In relies on Nvidia BlueField-4, Nvidia DOCA and Nvidia Spectrum-X Ethernet. BlueField-4 is Nvidia's fourth-generation data processing unit, a networking platform engineered for agentic AI factories and large-scale infrastructure requirements. According to Shainer, BlueField plays a critical role in managing agentic AI security.

"You can think about BlueField or Blue-Field-4 as a device that has in-silicon security engines that are working out of band and collecting telemetry from everything, from memory, from storage, from GPUs, from CPUs," Shainer told theCUBE. "That's why that infrastructure connects to all other infrastructures and ensures security all around. That out-of-band monitoring enables BlueField to actually understand exactly what agents are doing."

Nvidia DOCA, a software platform, also powers Scale-In implementations. The platform accelerates the development of applications and services that offload, accelerate and isolate data center workloads. Developers using DOCA can build software-defined, cloud-native solutions with DPU and SuperNIC acceleration. Similar to how Nvidia CUDA provides a programming interface for GPU workloads, DOCA grants developers access to DPU acceleration capabilities.

"DPU brings the in-silicon engines security engines, storage accelerations," Shainer noted. "DOCA is the framework, the software, the APIs that enable you to access all of those engines that BlueField-4 provides."

Nvidia has strengthened agentic AI security through OpenShell, an open-source runtime designed to govern agent execution, access and inference routing. The company paired OpenShell with BlueField-4 to create a monitoring device that operates independently of agents, collecting information outside the GPU in an environment no agent can manipulate.

"I want to also have another device which is completely separated, completely isolated from the agents, that can monitor everything out-of-band," Shainer said. "That's what we're doing with BlueField-4, bringing all the security elements together with OpenShell, and now you have a secure infrastructure for agents. This is how the combination of OpenShell and BlueField-4 brings all the security elements that we need for gen AI."

The OpenShell Reference Architecture, released more recently, builds on this foundation by incorporating Nvidia Sentry, which operates on BlueField-4 using DOCA software. Sentry continuously observes agent activity and can isolate agents that attempt to exceed their defined software boundaries. This combination extends OpenShell's runtime controls with an independent enforcement mechanism.

https://www.youtube.com/embed/0eeJZiZvj64?feature=oembed