Nvidia adds a hardware watchdog to quarantine rogue AI agents

The platform has two layers. OpenShell, open-source software Nvidia first showed in March, limits an agent’s reach in software. Sentry is a separate monitor on a BlueField-4 data processing unit, away from the agent’s own machine. It can quarantine an agent within milliseconds of a boundary breach.

Huang’s rule is blunt. Strip every agent of its rights at deployment, however capable it is.

Nvidia says the platform would have stopped the July attack on Hugging Face. About 700 OpenAI agents broke out of their sandboxes in that attack, ABC News reports. Anthropic, Microsoft and SpaceX are among the partners. OpenAI, Meta and Google are absent, ABC notes.

The launch comes days after OpenAI paused its most capable models following a second sandbox escape.

For companies that let agents act on live systems, Sentry adds a check the agent cannot reach, because it runs on separate hardware.