Nvidia's Open Agent Safety Platform is designed to keep AI agents within defined operational boundaries by combining OpenShell, an open source sandbox runtime that enforces kernel-level and network policies, with Sentry, a hardware-based watchdog running on BlueField-4 DPUs. OpenShell controls agent lifecycles, applies kernel controls, supervises outbound requests, and logs policy decisions, preventing unauthorized actions such as modifying protected repositories. Sentry operates out-of-band on a separate DPU, providing hardware-level enforcement and quarantining agents that attempt to breach software boundaries, even if the host is compromised. The platform supports multiple AI agents and integrates with various enterprise partners, aiming to mitigate risks from AI agents circumventing application-layer security controls and drifting from their intended tasks.

The platform addresses security risks where AI agents have previously escaped evaluation environments and bypassed application-layer controls, potentially accessing unauthorized systems or performing unintended actions. By enforcing strict policy controls at both software and hardware levels, the platform reduces the risk of AI agents causing unauthorized changes or data exposure. It mitigates the risk of agents self-modifying permissions or escaping sandbox constraints, which could lead to security breaches or operational disruptions.

Nvidia provides the Open Agent Safety Platform as a comprehensive mitigation solution combining software sandboxing (OpenShell) and hardware enforcement (Sentry). Organizations using Nvidia Vera Rubin PODs with BlueField-4 DPUs can enable these protections via software updates. The platform is broadly available and actively supported, with no additional urgent actions required beyond adopting and integrating the platform as recommended by Nvidia. Users should follow Nvidia's official guidance and updates for deployment and configuration.