NVIDIA's Vera Rubin platform offers a robust suite of security features designed to protect the integrity, confidentiality, and availability of AI workloads, particularly those involving complex agentic AI. A cornerstone of its security architecture is the third generation of NVIDIA Confidential Computing, which provides a unified, trusted execution environment across the entire rack-scale system. This extends across all 36 NVIDIA Vera CPUs, 72 NVIDIA Rubin GPUs, and the NVLink fabric connecting them, safeguarding data and models not only at rest and in transit but crucially, also while in use during computation. This comprehensive approach ensures that proprietary models, training data, and inference workloads are protected from unauthorized access and tampering throughout their lifecycle, complemented by attestation services that offer cryptographic proof of compliance.
Further enhancing the platform's security posture are specialized hardware and software components. The Vera Rubin NVL72 establishes a unified security domain that isolates GPU execution, memory, and register states, preventing leakage or compromise of sensitive AI assets. Additionally, the NVIDIA BlueField-4 Data Processing Units (DPUs) play a critical role by offloading and accelerating security tasks such as networking, storage encryption, virtual switching, and telemetry. These DPUs, running NVIDIA's DOCA software framework, provide an isolated and secure infrastructure layer, freeing up CPU and GPU resources while bolstering the overall security framework. This integrated hardware-software co-design ensures that demanding AI workloads can run securely at scale, even in shared or cloud environments, without compromising performance.
For developers building agentic AI applications, the Vera Rubin platform incorporates software-level security guardrails through components like OpenShell and NemoClaw. OpenShell is an open-source runtime that enforces crucial security and privacy policies for autonomous agents, acting as a sandbox layer to control agent actions and data access. NVIDIA's NemoClaw, an open-source stack built upon OpenClaw, integrates these privacy and security controls. The open-source nature of OpenShell allows security experts and developers to inspect, verify, and enhance its capabilities, fostering transparency and trust in the agentic AI ecosystem. This combination of hardware-level confidential computing and software-defined security mechanisms provides developers with a robust and transparent foundation for creating, deploying, and managing secure agentic AI systems.