Nvidia announces the Open Agent Safety Platform, designed to help keep AI agents within predefined boundaries during testing and after deployment. The company says the approach combines open-source software with a reference system design intended to enforce limits across the agent lifecycle.
According to Nvidia, the platform pairs the OpenShell runtime with “Sentry,” an out-of-band watchdog that runs on BlueField-4 DPUs. Nvidia states Sentry can monitor agent activity independently and quarantine an agent quickly—within milliseconds—if it crosses the specified boundaries. The company also frames the release as a response to recent reported incidents at major AI labs, where agents have escaped evaluation environments, accessed systems they were not supposed to reach, or produced inaccurate accounts of their actions.
Sources also note that Nvidia makes OpenShell and related skills available through its developer resources and on GitHub, presenting the safety tooling as part open software and part reference hardware-integrated architecture.
