NVIDIA unveils Open Agent Safety Platform to police AI agents

NVIDIA's pitch rests on a simple division of labor: OpenShell watches and constrains the agent where it runs, on the NVIDIA Vera CPU, while Sentry sits on the NVIDIA BlueField-4 DPU and cuts off agents that try to leave the permitted boundary. That combination is what NVIDIA describes as layered software and hardware security, and it addresses a specific worry — that an agent's own environment is not a safe place to put the control system.

The five principles NVIDIA lays out go further than the two products. Two of them are structural rather than technical: monitoring should happen out of band, meaning the agent need not know it is being watched, and control over the path to the model should yield both a good observation point and a kill switch. A third calls for verifiability before execution, so a policy can be shown not to escape the operator's intent. The last is about openness — agent runtimes and policy languages should be open so any provider can plug in, an argument that layers are owned separately by labs, companies and hardware providers.

Jensen Huang frames the whole thing around trust rather than raw capability, saying the most trusted AI must be built alongside the most capable. Whether that framing translates into adoption beyond the 100-plus industry partners NVIDIA cites likely depends on how readily third parties embrace the open runtime and policy language, and on whether agents can be governed without slowing the work they are meant to do.