Nvidia announced an AI-agent safety platform on Sept. 28 that pairs OpenShell, an open-source runtime for restricting an agent’s access, with Sentry, an optional hardware-monitoring design for BlueField-4 data-processing units. OpenShell is broadly available; Sentry is a reference design for an additional layer outside the agent’s host computer.
OpenShell places an agent in a sandbox and applies operator rules to its files, processes, network connections, tools and credentials. Requests pass through a supervisor, which can restrict an integration to read-only access or require a person to approve a request for broader permissions. Credentials remain outside the agent’s workload and are supplied only for permitted destinations. A formal checker verifies the access rules operators set; it does not prove that the goal of the agent or every permitted action is safe.
Sentry is meant to add a separate enforcement point. It runs on a BlueField-4 data-processing unit outside the computer hosting the agent. Nvidia says BlueField-4 sits on the only path from a compute server to the model in its planned Vera Rubin POD systems, giving Sentry a way to inspect and interrupt agent traffic outside the host. The company says it can quarantine an agent in milliseconds, but the launch materials do not provide an independent test or timing method for that claim.
The software and hardware layers have different availability. OpenShell is open source under Apache 2.0, and Nvidia says it can be extended to third-party compute. Sentry is an optional reference design tied to BlueField-4. Organizations can therefore try the runtime controls without deploying the hardware layer.
Platform coverage depends on the permissions operators define and the traffic routed through Sentry.
Sources
Photo: CSIRO, CC BY 3.0, via Wikimedia Commons. The image is illustrative and does not show Nvidia hardware.
