NVIDIA Debuts Open Agent Safety Platform to Control Autonomous AI Agents

# What NVIDIA announced

On September 28, NVIDIA unveiled the Open Agent Safety Platform, a two-part safety system designed to contain and control autonomous AI agents. The platform pairs an open-source runtime, OpenShell, with a hardware-based monitor called Sentry that runs on NVIDIA BlueField-4 data processing units (DPUs).

# Why this matters now

# together

OpenShell is a secure runtime that establishes a safety zone around an agent while it runs. It traces agent actions, enforces rules about what the agent can touch or do, and records activity. NVIDIA says OpenShell runs on its Vera CPUs and can also operate on systems using Arm and Intel processors.

Sentry runs separately on BlueField-4 DPUs as an independent observer. It continuously monitors agent behavior and can quarantine and stop an agent within milliseconds if the agent attempts to move outside permitted boundaries. Because Sentry sits on a DPU, it provides monitoring below the regular compute and application layers.

# Who's involved

NVIDIA reports more than 100 organisations are already working with its safety tools. Publicly named collaborators include Anthropic, Microsoft, Salesforce, SAP, Scale AI, ServiceNow, Cisco, CrowdStrike and Palantir. NVIDIA also says it expects the platform to be useful beyond enterprise software, for example in robotics and self-driving systems.

# Availability and ecosystem approach

OpenShell and related tools are available through NVIDIA's developer resources and GitHub. NVIDIA is also partnering with industry groups and government agencies to develop standards and best practices for deploying AI agents with these infrastructure-level controls.

# What the platform does not claim

NVIDIA positions the platform as a way to monitor, limit and stop agents at runtime. The announcement does not claim to change how models are trained, nor does it replace application-level governance or organizational policies. The emphasis is on adding a second line of defence at the runtime and hardware levels to reduce the risk of agents bypassing existing controls.

# Practical implications for adopters

Adopting Open Agent Safety Platform involves integrating OpenShell into the agent runtime and deploying Sentry-capable DPUs in the compute path where independent monitoring is required. Organisations running agentic systems that must access internal systems, files or external tools can use these components to enforce stricter boundaries and produce audit trails of agent activity.

# Short summary

NVIDIA's Open Agent Safety Platform combines OpenShell (a runtime boundary and policy enforcer) with Sentry (a DPU-based watchdog) to detect and stop agents that try to exceed their permissions. The tools are open-source or available via NVIDIA developer resources, already have industry collaborators, and target use cases that include enterprise applications, robotics and self-driving systems.