Nvidia Rolls Out New AI Safety Platform with Hardware-Level Watchdog

Nvidia has announced a new software tool aimed at containing runaway AI agents that can cause harm to their environments. The Open Agent Safety Platform, unveiled by Nvidia on Monday, includes two key elements: OpenShell and Sentry.

OpenShell is a 'sandbox' for AI agents, where they operate within a restricted workspace with a set of rules to follow. This approach prioritizes technical restrictions over trusting the agent to adhere to written instructions. According to Justin Boitano, Nvidia's vice president of enterprise AI, 'Agents can drift when instructions are ambiguous... An agent cannot be expected to fully police its own behavior.' OpenShell provides a 'secure runtime boundary' that traces all actions and enforces policy as agents run on Nvidia's Vera chips.

The second element, Sentry, operates at the hardware level as a 'watchdog' that monitors AI agent behavior. If an agent attempts to act outside set boundaries, Sentry can quarantine it instantly. Nvidia describes Sentry as a security checkpoint separate from the agent and computing system, acting as a backstop.