NVIDIA sets out Open Agent Safety Platform for AI agents
NVIDIA announced an Open Agent Safety Platform on 28 September, bringing together software controls and a reference hardware design intended to keep autonomous AI agents within defined boundaries. The announcement addresses a practical enterprise concern: an agent that can use tools and data may also take an action outside the permissions its operator intended.
What is in the platform
The platform has two main parts. OpenShell is an open-source runtime that traces an agent’s actions and enforces policies while it operates. NVIDIA says it can work with different AI models and can be extended to third-party compute platforms, including Arm and Intel systems. OpenShell is broadly available, according to the company.
Sentry is a separate reference design for a watchdog on NVIDIA BlueField-4 data processing units. It monitors behavior outside the agent’s own software environment and is designed to quarantine an agent that crosses a boundary. NVIDIA describes millisecond response, but that is a company claim about the design rather than an independent performance finding.
Why it matters
An AI agent differs from a chatbot that only answers questions. It may call applications, access files or services, and continue a task through multiple steps. A prompt or application-level instruction alone may not be enough to limit what it can do if the surrounding tools are too permissive. The proposed approach puts enforcement in the runtime and an isolated hardware layer, so the boundary need not rely solely on the agent following instructions.
That does not make an agent automatically safe. Organisations still have to define appropriate permissions, monitor what tools an agent can reach, test failure cases and decide when a person must approve an action. The effectiveness of Sentry in a particular deployment will depend on the implementation and its policies.
Who is involved
NVIDIA names a broad group of collaborators across AI, security, cloud and enterprise software. It says OpenShell software and related developer resources are available now. The Sentry component is described as a reference system design, so readers should distinguish the available software from a fully deployed security system in their own environment.
What to watch
The useful questions for prospective users are which actions are logged, where a policy is enforced, how quickly an agent can be stopped, and what happens if the monitoring layer fails. Those details matter more than a broad promise of “safe agents.” The announcement signals a move toward controls outside the model as companies give agents more consequential work.
