Nvidia is building agent security into the networking layer of its AI factory stack. According to SiliconANGLE, the company has added Scale-In as a new pillar of its AI networking infrastructure, aimed at securing, managing and operating agentic AI factories. The stated goals are secure access, fast data movement, tenant isolation, simpler operations and more predictable performance as agents take on larger roles.
Gilad Shainer, Nvidia's senior vice president of networking, told theCUBE that agentic workloads differ from earlier use cases because many agents run on shared infrastructure at once. Each must be kept inside its sandbox, and protection has to extend to memory and storage as well as compute.
Scale-In runs on Nvidia BlueField-4, the company's fourth-generation data processing unit, with Nvidia DOCA software, and connects over Spectrum-X Ethernet. Shainer described BlueField-4 as carrying in-silicon security engines that operate out of band and gather telemetry from memory, storage, GPUs and CPUs. DOCA is the software framework and set of APIs that lets developers reach those engines, much as CUDA does for GPUs.
The design rests on separation. Nvidia paired its open-source OpenShell runtime, which governs agent execution, access and inference routing, with BlueField-4 because it wanted a monitoring device that no agent could influence. Shainer said the aim is a device fully isolated from the agents that can observe everything out of band.
The approach has widened since August. In September, Nvidia introduced its Open Agent Safety Platform, combining OpenShell with Nvidia Sentry for out-of-band monitoring and hardware-based enforcement. An October announcement added the OpenShell Reference Architecture, which brings together Vera Rubin infrastructure, Vera CPUs and BlueField-4 for secure agent execution across enterprise systems. Per Nvidia, Sentry runs on BlueField-4 with DOCA, watches agent behavior continuously and can quarantine agents that step outside their defined software boundaries.
The coverage is based largely on Nvidia's own description and an interview with one of its executives. It offers no performance data, deployment details or customer names.
Why it matters: If agents become routine tenants of AI factories, enforcement that sits on separate hardware could become part of how operators think about isolation and trust. Whether this design delivers the predictable performance and containment Nvidia describes will depend on independent deployment evidence, which the source does not provide.
TwinStack automated analysis, based on reporting by SiliconANGLE. Check the original for full details.