NVIDIA places agent permissions outside the AI’s control

Its controls address files, services and credentials—the capabilities an agent can use to act on a person’s behalf.

The Open Agent Safety Platform combines OpenShell, an open-source runtime, with an optional hardware security layer called Sentry. OpenShell 0.1.0 places the agent inside a sandbox and checks its access from outside the workload. A team can permit a particular data query while withholding permission to change the same system. Credentials can remain outside the agent, attached only to authorized requests.

Sentry is a separate reference design running on NVIDIA BlueField-4 processors. NVIDIA says it can monitor activity independently and quarantine agents in milliseconds. That is the producer's claim, not a security result we reproduced. The software and hardware layers should also be distinguished: trying OpenShell does not mean buying the complete NVIDIA reference system.

The useful opportunity is to make permissions inspectable before handing an agent consequential work. NVIDIA's policy prover compares a candidate policy with an allowed boundary and can show an example of access that exceeds it. Its guarantees cover the behavior represented in its model; they are not a proof that an agent, application or deployment is safe.

For an engineering team, begin with the documented policy walkthrough, which demonstrates requests without needing an API key or language model. Test the permitted and forbidden operations on disposable data, inspect the resulting records, and check the platform requirements before considering production. A clear boundary is useful only when it matches the work you actually intend to authorize.