NVIDIA Introduces Open Agent Safety Platform for AI Agent Isolation
NVIDIA introduced the Open Agent Safety Platform, combining the open-source OpenShell runtime with optional Sentry hardware oversight for BlueField-4.
AI agent security is the focus of the new Open Agent Safety Platform, which NVIDIA announced on September 28, 2026. The solution combines the open-source OpenShell runtime software with NVIDIA’s Sentry reference design. Its goal is to manage AI agents’ activity, isolate them, and enforce rules for their access and operations.
The announcement does not concern a specific vulnerability or confirmed exploitation. NVIDIA is responding to security risks associated with agents that may execute code, work with data, and use external tools. In such use cases, setting rules only within the application is insufficient if an agent can exceed its assigned permissions or bypass application restrictions.
OpenShell Creates a Security Boundary for AI Agents
OpenShell forms the software component of the platform. According to NVIDIA, the runtime creates a boundary around the agent, monitors its actions, and enforces defined policies. NVIDIA says OpenShell is now generally available. SecurityWeek independently confirmed the basic architecture and identified the OpenShell release as version 0.1.0; it also describes it as a mechanism for isolating agents.
The software is not intended to be limited to a single hardware architecture. NVIDIA says OpenShell can be extended for Arm and Intel computing platforms. This differs from the hardware component of the design, which is tied to NVIDIA’s specific ecosystem.
Sentry Adds Oversight Outside the Agent Runtime
The second component is NVIDIA Sentry, an optional out-of-band watchdog designed for the BlueField-4 DPU. It is intended to independently monitor agent activity and enforce policies outside the runtime environment itself. This is a reference design, not a universally described layer for arbitrary server hardware.
NVIDIA claims that Sentry can quarantine an agent within milliseconds. However, this claim was not independently tested in the available sources. The full hardware-oversight model also requires the BlueField-4 ecosystem, while OpenShell reportedly has broader potential support for computing platforms.
Why AI Agent Security Matters
For enterprises deploying autonomous agents, the ability to isolate them, control access, and audit activity is especially important. NVIDIA’s platform shifts the focus from rules stored only in the application to enforcing them through a separate oversight layer as well.
NVIDIA also says it is working with more than 100 organizations. The extent of actual production deployment, the availability of individual integrations, and the level of involvement of these organizations have not been independently confirmed. The company also notes that several of the described products and features are at different stages of development and that their availability may follow a future schedule.
What to Watch Next
- independent security testing of OpenShell and Sentry’s effectiveness,
- availability of Sentry beyond reference configurations with BlueField-4,
- confirmed production deployments and integrations reported by partners,
- OpenShell licenses, supported platforms, and the pace of future releases.
Sources
- NVIDIA Newsroom – Confirms the announcement date, the platform’s composition of OpenShell and Sentry, the BlueField-4 architecture, and OpenShell’s general availability.
- SecurityWeek – Independently describes the platform’s two main components, OpenShell version 0.1.0, its isolation mechanisms, and the optional Sentry hardware layer.
Verified and updated: 09/28/2026 15:22
