NVIDIA Launches Open Agent Safety Platform with 100+ Enterprise Partners

NVIDIA launched its Open Agent Safety Platform on September 28, marking a significant shift in how the industry approaches AI agent containment. The platform pairs OpenShell, an Apache 2.0 open-source CPU runtime that isolates agent system calls, with Sentry, a hardware watchdog running independently on BlueField-4 chips.

Key points:

• Sentry operates outside the computer the AI is running on and can quarantine a misbehaving agent within milliseconds • Over 100 partners signed on at launch, including Microsoft, Anthropic, and Palantir • The platform moves containment from software policy into independent hardware that the AI itself cannot touch or persuade • Azure confirmed it will support the platform across its cloud infrastructure

This approach fundamentally changes the containment model. Until now, most AI guardrails were instructions baked into software — like a lock on a door. NVIDIA's approach adds a separate, independent camera and alarm system that doesn't trust the door's lock at all.

For enterprises running autonomous AI agents, this launch establishes a new baseline question for vendor evaluation: is your AI watched by hardware, not just software?

Why It Matters: Hardware-level containment represents a fundamental shift in AI safety architecture. Expect 'is this watched by hardware, not just software' to become a standard enterprise procurement question within the next two quarters.