NVIDIA described its Open Agent Safety Platform on September 28 as a layered reference design for controlling AI agents. It combines OpenShell, an open-source runtime that places agents in sandboxed environments, with an optional NVIDIA Sentry layer on BlueField-4 data processing units. NVIDIA says the two layers can apply policies outside the agent itself, with hardware monitoring meant to provide an independent observation and enforcement point. These are the company's architecture and engineering claims, not an independent security audit.
The immediate software component is OpenShell 0.1.0. NVIDIA says it can restrict which services, files, credentials, and operations an agent can access without requiring developers to rewrite the agent. Its September 28 technical walkthrough describes a gateway, supervisor, sandbox, policy checks, and a formal policy prover. OpenShell is not the same thing as Sentry hardware monitoring: NVIDIA describes the BlueField-4 layer as optional, and the more complete Vera Rubin POD design depends on specific infrastructure.
The distinction matters for teams evaluating autonomous coding agents. A published reference architecture does not establish that every enterprise has deployed it, that all agent behavior is captured, or that a sandbox cannot be escaped. NVIDIA's documentation explains how it intends the controls to work, while real-world protection depends on configuration, coverage, and testing. The next useful evidence will be independent assessments and details of deployed systems, not a blanket claim of safety.