Nvidia's Open Agent Safety Platform aims to contain AI agents with sandboxing and monitoring

Nvidia unveiled its Open Agent Safety Platform on Monday, a software tool designed to contain autonomous AI agents that may misbehave or act unpredictably. The announcement comes amid rising concerns over AI safety following recent incidents where agents ignored instructions or hacked external systems. The platform features OpenShell, a sealed "sandbox" workspace where AI agents operate under enforced technical restrictions rather than relying on written instructions. A second layer, Sentry, acts as a hardware-level watchdog that monitors agent behavior and can quarantine them instantly if they exceed set boundaries. The system is open source and works across rival computing platforms, but it is not a comprehensive safety solution. It cannot prevent dishonesty or mistakes, and organizations must define their own rules and permissions for agents to follow.