Nvidia puts AI agents behind new controls. The promise still needs testing

Its safety platform separates enforcement from the model, with software boundaries and a hardware watchdog designed to stop unauthorized actions.

Nvidia introduced a new platform on September 28 intended to constrain autonomous AI agents, arguing that systems able to take action need security controls outside the models directing them.

The Open Agent Safety Platform combines OpenShell software with a reference design called Sentry. Nvidia says OpenShell controls the environment in which an agent executes tasks, while Sentry monitors behavior independently on BlueField-4 hardware. The company says that second layer can quarantine an agent attempting to escape its boundaries within milliseconds.

Those are vendor claims, not the findings of an independent security assessment. The announcement establishes what Nvidia is offering and the architecture it proposes; it does not establish that every attack or unauthorized action will be stopped in every deployment.

The approach addresses a practical change in how AI is used. An agent can read and write files, invoke software and contact external services. A mistaken answer is therefore only one possible failure. A system carrying out instructions can also make a real change to data or infrastructure, which makes permissions and enforcement central operational questions.

HPE, a launch partner, offered more detail about its implementation. It said OpenShell uses isolated environments to govern reading, writing, execution and network access, with controls operating outside the agent process. When an agent encounters a policy limit, operators retain authority to approve a change. HPE plans availability of its Private Cloud AI integration in the fourth quarter of 2026.

That timetable matters. Open-source software availability and a finished integration in a customer’s infrastructure are different milestones. Organizations must still connect controls to their own identities, policies and audit arrangements. HPE’s announcement describes a planned integration, rather than evidence that every advertised capability is already operating for all customers.

Nvidia also says the software can be extended beyond its own compute platforms, including to Arm and Intel systems. That offers a route toward broader use, although the hardware watchdog described in the launch relies on Nvidia components. Buyers will need to distinguish which protections are available in the software alone and which require the reference hardware.

For an enterprise evaluating the platform, the relevant controls are the permissions granted before execution, enforcement during an action and records available afterward. Nvidia and its partners describe products for those tasks; the announcement provides no independent, cross-industry measurement of their effectiveness.