Nvidia Builds a Watchdog for AI Agents Into Its Chips

Nvidia has introduced Sentry, a hardware watchdog for data centers built on BlueField-4 chips. It cuts an AI agent off from the network if the agent exceeds its allowed access rights, and it does so within milliseconds. In September, a similar protective layer at the company already caught a violation. The alarm went off 12 minutes after the agent's first unauthorized outbound call, but fully shutting it down back then took two hours and forty-four minutes. Nvidia compares this work to how the browser once isolated every website from every other one, and in doing so made the internet safer.

From my own experience with agents in Claude Code, I grant an agent access for one task at a time rather than opening everything up at once. Nvidia is baking the same idea into data center hardware instead of leaving it up to whoever writes the prompt. For big GPU farms run by cloud providers, that's a sensible move. For an everyday vibe coder running an agent in their own terminal, a hardware watchdog does not add much. You still have to limit access yourself, with logs and separate keys.