Red Hat and Nvidia expand partnership to secure autonomous AI agents

The companies are targeting two risks: agents operating beyond their authorized limits, and attackers compromising agents through prompt injection or supply chain attacks. Their approach places security controls outside the agent and, where suitable, on the host operating system. John Fanelli, Nvidia's vice president of enterprise software, said agents should not enforce their own boundaries, and that those controls belong in the runtime and the infrastructure beneath it.

Nvidia's OpenShell provides a sandbox runtime for agents, and Red Hat says it is working to make its reference architecture a native capability of its AI platforms. Below that layer, Nvidia's Sentry framework runs on BlueField data processing units and uses DOCA Argus to monitor workloads from a separate hardware security domain without installing monitoring software on the host. Red Hat says Sentry can quarantine a compromised agent in milliseconds, and the design builds on BlueField-3 support in Red Hat OpenShift 4.22.

A joint reference design pairs Model-as-a-Service access controls and an AI Gateway with vLLM, llm-d, Nvidia Dynamo and Nvidia NIM microservices, adding self-service model discovery, access policies, rate limits and subscription tracking. The companies also added support for Nvidia Nemotron models, including Nemotron 3.5 Super VL, in vLLM. Red Hat Enterprise Linux for Nvidia now supports Vera CPUs and is prepared for the Vera Rubin platform. Cisco, Dell and Lenovo are listed as partners offering the capabilities, while native OpenShell integration is still in progress.

Teams running agents on Red Hat OpenShift or Nvidia hardware may be able to rely more on controls outside the agent process, such as sandboxing and host-level isolation, instead of trusting the agent to limit itself. A gateway layer with access policies, rate limits and subscription tracking could simplify quota and cost management across hybrid deployments. Several pieces are reference designs or plans tied to upcoming hardware, so check availability before building on them.