Nvidia Enhances AI Agent Security with Scale-In and Sentry

Nvidia combines hardware and software to manage agentic AI in data centers. Beyond a new network component called Scale-In, the company activates BlueField-4, DOCA, OpenShell, and Sentry to monitor out-of-band, isolate tenants, and, according to the company, quarantine agents that stray outside their software boundaries.

Sentry continuously monitors and can isolate deviant agents

Nvidia integrates Sentry into the OpenShell Reference Architecture, executing on BlueField-4 using DOCA software. According to the company, Sentry provides continuous observation of agent behavior. Nvidia also states that Sentry can quarantine agents attempting to breach their defined software limits. The addition of this independent application layer extends the runtime controls of OpenShell. This reference architecture, announced in October, also includes the Vera Rubin infrastructure, Vera CPUs, and BlueField-4 to secure execution at the enterprise system scale.

OpenShell and BlueField-4, a duo for out-of-band control

OpenShell, Nvidia's open-source runtime, governs the execution of agents as well as access and routing of inferences. Nvidia has paired it with BlueField-4 to collect signals outside the GPU and maintain a monitoring zone unaffected by agents. Gilad Shainer describes the need for a separate and isolated device capable of monitoring everything out-of-band, a role fulfilled by BlueField-4 combined with OpenShell. According to him, this combination brings together the necessary security elements for generative AI. He specifies that BlueField-4 includes security engines operating out-of-band and capable of aggregating telemetry from memory, storage, GPUs, and CPUs to accurately characterize agent activity.

Scale-In targets the AI factory with isolation and data throughput

Scale-In provides acceleration specifically designed to secure, manage, and operate agentic AI factories. The design aims for secure access, high-throughput data transfers, tenant isolation, simplified operations, and more stable performance. The stated goal is to enhance security at all levels of the infrastructure. Nvidia positions Scale-In as a new pillar of its AI network. Gilad Shainer reminds us that agentic AI requires many agents to operate in parallel, which necessitates strengthening safeguards to contain them and prevent unwanted behaviors, including at the memory and storage levels. He detailed these ambitions during an interview in a series dedicated to AI factories.

BlueField-4, DOCA, and Spectrum-X at the heart of the architecture

Scale-In relies on the BlueField-4 DPU and DOCA software, interconnected via Spectrum-X Ethernet. BlueField-4 is Nvidia's fourth generation of data processors, designed for the networking needs of AI factories and large-scale infrastructures. Gilad Shainer considers BlueField a cornerstone of agent security management. DOCA powers deployment and accelerates the development of applications capable of offloading, accelerating, and isolating critical workloads. Developers can build software-defined cloud-native solutions accelerated by DPU and SuperNIC. DOCA provides access to DPU accelerations, similar to CUDA for GPUs, and offers, according to Shainer, the framework and APIs to utilize integrated security and storage engines.

An extended security strategy in three phases

Introduced in August, Nvidia's approach has rapidly expanded. In September, the Open Agent Safety Platform combined OpenShell and Nvidia Sentry for out-of-band monitoring and a security application rooted in hardware. In October, an OpenShell reference architecture was presented to unify Vera Rubin, Vera CPUs, and BlueField-4 around the secure execution of agents across enterprise systems.