Breaking

NVIDIA Introduces Open Agent Safety Platform With OpenShell Runtime and Sentry Watchdog

September 28, 2026 at 08:02 EDT

AI Agents
Security
Infra & Chips

NVIDIA announced NVIDIA Open Agent Safety Platform on Sept. 28, a security stack meant to keep autonomous AI agents inside enforceable boundaries from testing through deployment. The system pairs NVIDIA OpenShell, an open-source runtime policy layer, with NVIDIA Sentry, a BlueField-4 DPU-based watchdog designed to quarantine agents that move outside their boundaries in milliseconds.

Sept. 28, 2026 · NVIDIA

Agent safety moves below the model

NVIDIA Open Agent Safety Platform pairs OpenShell runtime policy with Sentry monitoring on BlueField-4 DPUs, aiming to keep long-running agents within enforceable software and hardware boundaries.

Core platform layers

100+

Organizations working with the platform

120+

Open Secure AI Alliance participants

Arm Neoverse V2 cores in BlueField-4

The ecosystem comparison

Each block represents about 10 organizations.

100+

Platform organizations

120+

Alliance organizations

How the control loop works

Run

Agent executes in a sandbox

→

Check

Policy gates files, tools and networks

→

Escalate

Human approval can grant scoped access

→

Quarantine

Sentry can stop out-of-bound behavior

Why teams like it

Controls sit below prompts and agent harnesses, so policies can keep applying while agents call tools, hit errors and retry over long workflows.

What limits adoption

The software runtime can extend across compute platforms, but the independent Sentry enforcement path depends on BlueField-4 DPUs.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.