HPE brings NVIDIA OpenShell agent governance to Private Cloud AI

Controls enforced outside the agent

HPE frames agent governance around three operational questions: where agents are running, what they can reach, and what they did. In its view, a policy document does not answer them. Controls enforced in the operating environment do.

NVIDIA OpenShell is the secure runtime at the center of the NVIDIA Open Agent Safety Platform, and it is now generally available. According to HPE, it is open source software that places controls around the environment in which an agent operates. It uses isolated sandboxes and policy enforcement to govern what an agent can read, write, execute, and access over the network, including where inference requests go.

These controls operate outside the agent process, which separates the control boundary from the model, agent harness, or application it governs. If an agent hits a policy limit, it can request a change, while operators keep approval authority.

The general availability release adds multi-tenancy, Kubernetes-based high availability, extensibility, centralized observability, and policy automation. HPE says enterprises can support different agents, models, and frameworks under one consistent runtime boundary.

What HPE Private Cloud AI adds

HPE Private Cloud AI is a turnkey AI factory solution within the HPE AI Factory with NVIDIA portfolio. HPE is integrating OpenShell into an operating environment designed to connect agent execution with enterprise identity, policy, approvals, observability, and audit.

The company says this gives platform and security teams one place to govern agent activity without forcing development teams onto a single agent framework. OpenShell supports a variety of models, harnesses, and agents across cloud, hybrid, on-premises, and air-gapped infrastructure. IT administrators control what agents can access, while developers keep flexibility in how they build them.

The stated objective is to give agents the access they need for approved work without unrestricted access to hosts, networks, credentials, data, or model endpoints.

Infrastructure-level protection

HPE and NVIDIA also describe a second layer below the runtime. NVIDIA Sentry is an independent watchdog that monitors agent activity and enforces security policies out of band. It runs on NVIDIA BlueField-4 DPUs and uses NVIDIA DOCA for hardware-isolated monitoring, agent governance, and policy enforcement.

• HPE plans to support BlueField-4 DPUs across its servers, AI rack-scale systems, HPE Private Cloud AI, HPE AI Factory at-scale, and HPE Sovereign AI Factory, with availability based on product lead times.
• OpenShell and its agentic sandboxes run on NVIDIA Vera CPU systems, including HPE ProLiant Compute DL394 Gen12, as well as other compute platforms.
• HPE is integrating NVIDIA Confidential Computing with HPE AI Factory solutions through HPE Services, protecting models and private data during execution for on-premises and sovereign deployments. Availability is planned for Q4 2026.

According to HPE, these capabilities can help customers address requirements tied to the Cybersecurity Maturity Model Certification, the National Institute of Standards and Technology 800 series, Security Technical Implementation Guides, and Federal Information Processing Standards. The requirements addressed depend on configuration and deployment.

Sovereignty and what it means for buyers

For governments, public sector bodies, and regulated industries, HPE says governance also means knowing who controls the system, where data resides, and who may operate the environment. HPE Sovereign AI Factory extends the HPE AI Factory with NVIDIA portfolio for environments needing greater control over data, infrastructure, operations, and compliance. Paired with a governed agent runtime, HPE says it can help organizations run agents near sensitive data within defined boundaries.

For enterprises weighing agentic AI in production, the pitch is that governance should be enforceable, observable, and auditable rather than only documented. Most of the HPE-specific pieces are still planned for Q4 2026, so the practical picture will depend on how those integrations ship.