Nvidia Unveils AI Agent Safety Platform With Independent Hardware Monitoring

TL;DR

• Nvidia has introduced an Open Agent Safety Platform designed to keep AI agents within defined operating boundaries.
• Its OpenShell software controls agent access and records actions as agents work.
• A separate Sentry system monitors activity on Nvidia’s BlueField-4 data processing units.

Nvidia has launched a platform that places security controls around AI agents at both the software and hardware levels. The Open Agent Safety Platform, announced Monday, is intended to help organizations control what agents can access during testing and deployment.

The announcement follows reports of AI agents bypassing application-level controls while carrying out assigned tasks. Nvidia’s approach puts some enforcement outside the agent itself, so restrictions do not depend solely on the model following instructions.

OpenShell Sets the Software Boundary

The platform includes OpenShell, Nvidia’s open-source software for running agents within defined limits. It traces their actions and enforces policies governing how they execute tasks and access other systems. Nvidia says OpenShell is broadly available and can be extended to work with third-party computing platforms.

That software boundary is paired with Sentry, a monitoring system designed to run on Nvidia’s BlueField-4 data processing units. Because Sentry operates on a separate processor, Nvidia says it can observe agent activity independently and stop an agent that attempts to move beyond its permitted boundaries. Its claim that Sentry can quarantine agents in milliseconds is a stated capability of the platform, not evidence that it would have prevented every previously reported incident.

Nvidia Draws Support Across the AI Industry

Nvidia named Anthropic, Microsoft, Hugging Face, Salesforce and several other companies among organizations working with the platform’s technologies. It described the wider effort as involving more than 100 organizations across AI, infrastructure, finance and robotics.

The launch reflects a practical challenge for companies deploying agents: as those systems gain access to more tools and data, operators need ways to see what they do and enforce limits while they work. Nvidia is offering software controls backed by a separate hardware monitoring layer. How well that combination performs across different deployments will depend on its implementation and testing.