Open Agent Safety Platform
A security platform and reference system design to set safer boundaries for AI agent deployment. nvidianews.nvidia.com ↗
NVIDIA Open Agent Safety Platform, announced September 28, 2026, is a two-part offering: OpenShell, a free Apache 2.0 runtime sandbox, and Sentry, a hardware-level watchdog running on BlueField-4 DPUs. The software layer is genuinely open and cross-platform, but the full-stack value proposition is tightly coupled to NVIDIA's own silicon roadmap (Vera CPUs, BlueField-4 DPUs), creating a meaningful lock-in risk for enterprises that run mixed-vendor infrastructure. The 100-plus partner backer list is impressive for a week-old launch, but no independent benchmarks exist for the "milliseconds" quarantine claim, no general-availability date has been published, and the real cost center (policy authoring, DPU hardware procurement, and ongoing maintenance) is largely absent from NVIDIA's marketing. OpenAI is conspicuously absent from the backer list, and critics note that runtime governance only protects agents you already know about, leaving discovery and inventory as unsolved governance gaps.