Nvidia is pitching its Open Agent Safety Platform as a way to put tighter limits on AI agents, with controls designed to govern what they can access and contain them if they cross a boundary. The platform brings together OpenShell software and a separate monitoring system called Sentry.
The company’s proposal puts agent safety in the surrounding computing environment. Developers would define the resources an agent needs for a task, while separate software and hardware enforce those limits. Nvidia argues that safeguards inside a model are insufficient to control everything an agent can do.
For Nvidia CEO Jensen Huang, that fits an engineering approach to AI safety: improve the systems around increasingly capable models so they can be used with greater confidence. The platform is intended to make those restrictions part of how agents run, from testing through deployment.
How OpenShell and Sentry divide the work
Nvidia describes OpenShell as an open source runtime that sets boundaries around agents running on CPUs. Its role is to govern an agent’s interactions with its environment, with enforcement taking place outside the agent’s own process.
Under that design, access is denied unless a policy permits it. Nvidia says agents run in isolated sandboxes with restricted file access and no direct network access. Network requests pass through a supervisor that checks them against policy, and permission decisions are logged for review.
That makes the intended restriction concrete: an agent should only be able to reach resources its operator has authorized. Nvidia says OpenShell also checks proposed policies to determine whether they would exceed the permitted access boundary.
Sentry is the additional hardware layer. Nvidia describes it as an independent watchdog running on BlueField-4 data processing units, specialized chips used in infrastructure and networking. Its monitoring sits outside the host software running the agent, adding a separate point of enforcement.
The company says Sentry can isolate agents that attempt to exceed their boundaries within milliseconds. That is Nvidia’s performance claim; it does not establish that every unsafe action will be detected or every security incident prevented. The design aims to keep enforcement separate even when the host environment cannot be trusted.
A platform for partners to build on
Nvidia positions the broader platform as a reference system design that other companies can incorporate into their own products. OpenShell supplies the open source software foundation, while Sentry provides an optional additional layer for organizations using the supporting hardware.
The company lists Cisco, Microsoft, Oracle Cloud Infrastructure, CoreWeave, Dell, HPE and Lenovo among infrastructure partners supporting the technologies. It also says OpenShell can be extended to work with third-party computing platforms, including those from Arm and Intel.
Nvidia says its collaboration with Anthropic connects Claude Managed Agents with OpenShell and BlueField controls. The stated aim is to give enterprises tighter control over agent access through the sandboxes where work takes place.
The distinction running through the platform is between an agent’s ability to reason and its authority to act. Nvidia’s bet is that companies can give agents useful work while keeping their permissions narrow, with the surrounding infrastructure responsible for enforcing the boundary.
