Nvidia has unveiled the Open Agent Safety Platform, a governance framework designed to control how autonomous AI agents operate across software, hardware, and physical systems as enterprise deployments of agentic AI become more complex.
The platform includes OpenShell, now broadly available, which creates a secure runtime boundary that constrains what an autonomous agent can do during task execution.
The protection sits outside the model and its harness, giving organisations an enforceable limit independent of the AI software itself. Nvidia says OpenShell adds minimal overhead when running on Vera, its purpose-built CPU for agentic workloads. As open-source software, OpenShell can also be extended to third-party compute platforms from Arm and Intel.
The hardware enforcement layer comes through Nvidia Sentry, a watchdog system that runs on BlueField-4 DPUs and monitors agent behaviour out-of-band. If an agent attempts to move beyond its defined software boundary, Sentry can quarantine and stop it within milliseconds, providing what Nvidia describes as in-silicon security enforcement.
Organisations can deploy elements of the platform selectively, according to their own requirements, rather than adopting the full stack at once.
The launch reflects a growing concern among enterprises and regulators: as AI agents take on more consequential tasks across more systems, the guardrails must extend beyond the model itself into the infrastructure running beneath it. Nvidia is positioning the Open Agent Safety Platform as that infrastructure layer.
