NVIDIA and Red Hat add AI controls

Red Hat and NVIDIA said on May 12 they had expanded their joint AI software stack with new controls for autonomous agents, adding policy enforcement, sandboxing and audit features aimed at enterprise deployments. The updates were announced at Red Hat Summit in Atlanta as part of Red Hat AI Factory with NVIDIA, a co-engineered platform the companies launched in February for building and running AI applications across on-premises systems, public clouds and edge environments. Red Hat said the latest release is focused on “autonomous, long-running agents” and on giving IT teams tighter oversight as those systems move into production. NVIDIA documents describe the platform as a unified foundation for building, deploying and scaling AI workloads across hybrid cloud environments. ### What exactly did Red Hat and NVIDIA add this week? Red Hat said the new package centers on OpenShell, an open source project founded by NVIDIA that provides a sandboxed runtime for autonomous AI agents. In the company’s May 12 announcement, Red Hat said OpenShell is designed to govern how agents execute, what they can access and where inference is routed, while giving customers “a single, unified policy layer” to control and monitor agent behavior. (redhat.com) NVIDIA and Red Hat also tied the update to operational controls that enterprises usually ask for before broader deployment: stronger policy controls, safer tool use and auditable operations. Red Hat said joint engineering work is underway to integrate OpenShell with its full-stack AI platform, and said it is contributing to the upstream OpenShell project to help standardize management of autonomous agents across hybrid cloud environments. (redhat.com) ### Why is OpenShell the key piece here? Joe Fernandes, a Red Hat vice president, wrote in a March 16 company blog post that one gap in the current AI stack is a dedicated layer that can give agents access to tools and services while enforcing security and privacy controls. Fernandes said OpenShell was built for that role, with agent sandboxing, deny-by-default policy and privacy-preserving routing. (redhat.com) Red Hat said OpenShell runs within Kubernetes and can be deployed on Red Hat AI, which lets customers connect agents to self-hosted models, MCP tools and other AI services inside hybrid infrastructure. That matters because the companies are pitching the product not as a chatbot layer but as runtime software for agents that can operate over longer periods and interact with production systems. (redhat.com) ### How are they handling security beyond software policies? Red Hat said the platform now extends controls to the hardware layer through confidential computing. In the May 12 announcement, the company said users can run confidential containers with NVIDIA Confidential Computing as part of Red Hat OpenShift sandboxed containers, in what it described as a technology preview. (redhat.com) The May 12 release said those controls are intended to protect against runtime compromise of agents, including cases where another agent has already been breached. Red Hat said the setup is reinforced by a zero-trust architecture that includes SELinux, FIPS compliance and NVIDIA DOCA-based runtime protection. ### Where does this fit in the broader Red Hat-NVIDIA partnership? (redhat.com) Red Hat launched Red Hat AI Factory with NVIDIA on February 24 as a co-engineered platform combining Red Hat AI Enterprise with NVIDIA AI Enterprise. At the time, the companies said the offering was meant to provide an end-to-end software foundation for organizations deploying AI at scale, with support across on-premises, cloud and edge environments. (redhat.com) NVIDIA’s deployment guide says the platform includes validated blueprints, model-serving tools and developer environments intended to move teams from experimentation to production. That earlier launch framed the current governance update as an extension of the same platform rather than a separate product. ### What comes next for customers? Red Hat said joint engineering work to integrate OpenShell into its AI platform is still underway. (redhat.com) The companies are also positioning confidential containers with NVIDIA Confidential Computing as a technology preview rather than general availability, which means the next milestone for customers will be broader product support and deployment guidance inside OpenShift environments. (docs.nvidia.com) Red Hat Summit 2026 runs through May 14 in Atlanta, according to the company’s event agenda. Red Hat’s summit materials direct customers to the session catalog and agenda for product roadmaps, labs and follow-on details from Red Hat and partner participants including NVIDIA. (redhat.com 1) (redhat.com 2)