NVIDIA announced its Open Agent Safety Platform on 28 September 2026, describing a system intended to govern AI agents from testing through deployment. An Arm Newsroom technical post published the same day outlined possible roles for Arm processors and infrastructure processors in that design. NVIDIA’s platform combines two components: OpenShell software, which limits what an agent runtime can access and do, and the NVIDIA Sentry reference design, which monitors agent behaviour independently of the system. NVIDIA says OpenShell is open source and can be extended to third-party CPU platforms, including Arm and Intel. The announcement should therefore not be described as an Arm-exclusive product or exclusive partnership; Arm’s post explains Arm-based infrastructure options in the design.

In the example outlined by Arm, the CPU layer running agents is separated from the layer that performs control and monitoring. Sentry, running on an NVIDIA BlueField-4 data processing unit, is designed to monitor agent activity, enforce policies such as network and data access rules, and quarantine an agent that attempts to move outside its boundaries. Arm describes this BlueField-4 configuration as powered by an NVIDIA Grace processor with 64 Arm Neoverse V2 cores. OpenShell creates a runtime boundary around permitted access and actions; Sentry’s separate infrastructure plane is intended to keep controls from depending solely on privileges held by the agent software. This is an architectural description, not an independent test result for security effectiveness in production or a performance benchmark.

The design addresses an engineering question raised by autonomous software that runs for longer periods and accesses tools and services: the surface for mistakes or abuse grows. A model or agent proposing an action and a system authorising that request should have separate roles; access policies should be observable and revocable. NVIDIA says Sentry is a hardware-supported control layer on BlueField-4 that continuously monitors agents, while OpenShell applies runtime boundaries and helps stop agents that try to exceed policy. Whether the design works across different organisations depends on coordinated management of models, tools, user identities, application permissions and infrastructure configuration. The presence of a security component does not mean a system is resilient to every attack.

For businesses, the potential value is improved oversight by placing the processor hosting an agent and the security control observing it in separate trust domains. That brings implementation and operating requirements: additional infrastructure, policy management, review of false positives, identity checks and protection of event logs. The platform offers security controls; it does not guarantee model quality, business outcomes or regulatory compliance for every application. NVIDIA introduced Open Agent Safety Platform as open software and a reference system design and announced participation from companies across the sector. Arm’s contribution here is a technical account showing CPU options compatible with its architecture and the use of Arm-based DPUs. The announcement gives no new Arm chip sale, order value or forecast of future revenue.