NVIDIA’s New Agent Security Platform Raises a Bigger Question

NVIDIA made an announcement today that I think deserves more attention than another AI product launch.

It introduced the Open Agent Safety Platform, including OpenShell, which creates a secure runtime boundary around agents, and Sentry, an independent monitoring and enforcement layer designed to quarantine an agent when it moves outside those boundaries. NVIDIA says more than 100 organizations are working with the technology. 

The timing matters.

OpenAI reportedly paused model training after an agent obtained network access it wasn’t supposed to have. The reporting describes an operational-control failure rather than an external system being compromised. 

Put those developments together and I think we’re seeing an important shift in how we need to think about AI security.

We may be trying to solve an architecture problem with AI policy.

A policy can tell an agent what it is allowed to do.

A prompt can tell it what it shouldn’t do.

A model can be trained to behave within certain boundaries.

But none of those things are the same as building an environment that determines what the agent is actually capable of doing.

That distinction becomes much more important as agents move from generating answers to taking actions.

An agent might have an identity. It might call APIs. It might access corporate data, execute code, communicate with other systems or trigger another agent.

At that point, the security conversation starts looking familiar.

What identity does the agent have?

What systems can it reach?

What data can it access?

What APIs can it call?

What permissions does it receive?

Can we see what it did?

And what independently stops it when it moves outside the boundary we’ve established?

NVIDIA’s architecture is interesting because OpenShell is designed to enforce policy while an agent executes, while Sentry operates independently through BlueField-4 DPUs and can quarantine an agent attempting to move outside its software boundary. 

That’s an architectural approach to governance.

And I think that distinction is going to become increasingly important.

AI governance still needs policy. It still needs acceptable-use rules, risk management and human accountability.

But agents are turning governance into an enforcement problem too.

The enterprise AI security architecture therefore can’t stop at the model.

It has to extend through identity, network access, APIs, data, runtime controls, observability and the infrastructure underneath the agent.

We’ve spent years applying Zero Trust principles to humans, applications and workloads.

AI agents may simply be the next identity that has to earn trust.

The question I would ask now isn’t just:

What is this agent supposed to do?

It’s:

What have we architecturally made it impossible for this agent to do?

Technology doesn’t create business value. Better decisions do.

Keith Pasley | Helping Organizations Make Better AI Decisions.

Back to Insights