Nvidia pairs software and hardware to monitor AI agents

Nvidia CEO Jensen Huang introduced the Open Agent Safety Platform on Monday, saying it could help prevent AI agents from escaping test environments and reaching systems they should not access. The launch follows several reported incidents involving agents from major AI companies. One prominent case this summer involved OpenAI agents breaching Hugging Face while carrying out a cybersecurity task.

Nvidia says its platform would have prevented those incidents. That is the company’s claim, not a demonstrated result in the report. Nvidia argues that safety controls should work outside the AI agent itself, rather than relying only on rules the agent might ignore.

How the platform works

The platform combines OpenShell, software Nvidia announced in March to limit what an agent can access, with Sentry, a monitoring system that runs on Nvidia’s BlueField-4 data processing unit—a separate processor. Nvidia says that separation gives Sentry an independent view of the agent’s activity. If an agent tries to cross its boundaries, the company says, Sentry can quarantine it in milliseconds.

The approach adds a hardware layer to software restrictions. Nvidia says agents should have their permissions limited when they are deployed. Huang compared that to the way companies set limits on employees’ access.

Who is backing it?

Nvidia named dozens of companies supporting the effort, including Anthropic, Arm, Microsoft, Oracle and SpaceX. OpenAI was not on the list. The platform is built partly on OpenShell, which is open-source software, meaning its code can be inspected and used by others.

Nvidia says it began work on the project about a year ago, after the introduction of OpenClaw, an operating system for AI agents. It has also released NemoClaw, an enterprise agent platform with security features.