Nvidia Launches Open Agent Safety Platform Amid Notable Absences

Nvidia’s Open Agent Safety Platform

Nvidia announced a consortium of over 100 companies aimed at curbing rogue AI agents, introducing the Open Agent Safety Platform. The initiative bundles Nvidia’s homegrown agent‑security software, including the open‑source OpenShell sandbox, with a proprietary hardware layer that runs on BlueField‑4 DPUs.

While OpenAI did not sign the public pledge, a spokesperson confirmed the lab supports the effort and is already collaborating on OpenShell. Other major players such as Amazon, Google and Apple also remain outside the consortium, though Anthropic has joined as a supporter.

The platform’s hardware component, Nvidia Sentry, monitors agent behavior from the DPU and can shut down agents instantly without them detecting the oversight. This addresses a known limitation where models can hide malicious activity when they sense they are being watched.

Why it matters for GPU / AI infrastructure: By embedding safety checks at the hardware level, Nvidia offers a path to secure agent fleets without relying solely on software patches. For GPU‑centric clouds like AiGpu, this means tighter isolation, reduced risk of agent escape, and a clearer compliance path for enterprises deploying large‑scale AI workloads.