Nvidia's New AI Watchdog Stops Rogue Agents Before They Cause Trouble

Nvidia announced a new system on September 28 designed to keep AI agents — software that can take actions on your behalf, like sending emails or booking flights — from doing things they aren't supposed to. The platform, called Open Agent Safety Platform, combines two pieces: OpenShell, free software that sets strict boundaries around each agent, and Sentry, a hardware design that watches agents from a separate chip. Think of it like a babysitter who can see everything the kid does — and pull the plug if things go wrong.

Here's the problem it solves. AI agents increasingly have access to your email, files, passwords, and company systems. If an agent gets confused or hacked, it could reach data it was never meant to touch. Nvidia's OpenShell lets companies define exactly which files, networks, and tools each agent can use. A supervisor sitting outside the agent checks every request against those rules. Nvidia says teams can apply these limits to agents they already use without rewriting them.

The second layer, Sentry, is more unusual. Instead of trusting the agent's own software to police itself, Sentry runs on separate hardware (Nvidia's BlueField-4 chips) and can quarantine an agent within milliseconds if it crosses a line. It also logs every action into a record for later investigation, connecting what the agent did, which rule applied, and which tool it used.

The catch: OpenShell is available now and open source, but Sentry is only a reference design — no price, no release date, and no proof yet of how it performs on real customer workloads. Nvidia says over 100 organizations are working with the platform, but that doesn't mean they've deployed both layers. For now, the practical step for companies is to map what each AI agent can reach, then test whether these controls actually hold up.