AI safety is moving out of policy documents and into the infrastructure itself. NVIDIA has launched the Open Agent Safety Platform, combining an open-source runtime boundary called OpenShell with a separate hardware-based watchdog, Sentry.