Nvidia thinks it can stop rogue AI agents without slowing AI down.

And its answer to the AI safety debate is basically: don’t regulate the intelligence. Build a better security layer around it.

Nvidia has introduced the Open Agent Safety Platform, a two-layer security system designed to control what autonomous AI agents can access and immediately isolate them when they cross established boundaries.

The company says the technology could have prevented the recent Hugging Face breach involving OpenAI models.

Here’s the real story:

The first layer, OpenShell, lets developers establish rules governing what an AI agent can access and then enforces those rules in real time.

The second, Nvidia Sentry, independently watches the agent for suspicious behavior and can quarantine it within milliseconds if something goes wrong.

That’s an important distinction.

Instead of assuming increasingly autonomous AI will always behave correctly, Nvidia’s approach assumes agents eventually will make mistakes and builds containment directly into the infrastructure running them.

It also offers a very different answer to the growing AI safety debate.

Some AI leaders are calling for slower frontier development and greater coordination. Nvidia CEO Jensen Huang has argued that safety is fundamentally an engineering problem and that advanced AI should continue moving forward while being rigorously tested and contained.

The next big AI safety breakthrough might not be teaching agents to never cross the line. It might be building infrastructure that physically stops them when they do.