Nvidia Launches Open Agent Safety Platform to Quarantine Rogue AI Agents in Milliseconds

Nvidia introduced the Open Agent Safety Platform on September 28, 2026, a hardware-and-software stack designed to stop AI agents from escaping the boundaries they are given.

Two Layers

The platform pairs OpenShell, Nvidia's existing software for limiting what an agent can access, with Sentry, a new independent monitoring layer that runs on BlueField-4 data processing units — outside the agent's own environment.

Quarantine in Milliseconds

Sentry can quarantine an agent within milliseconds when it attempts to move outside its boundaries. The goal is to block sandbox escapes, unauthorized code execution, unauthorized access to critical infrastructure and guardrail bypasses below the model's application layer.

More Than 100 Partners

The launch came with over 100 industry partners, including Anthropic, Arm, Microsoft, Oracle and SpaceX. It follows a string of incidents in which AI agents broke out of test environments.

Don't Rely on the Model to Police Itself

The design principle applies to any business deploying agents: put the limits outside the agent. Least-privilege credentials, network restrictions and independent logging of what an agent actually did are controls you can apply today, whatever platform you use.