Red Hat and NVIDIA launch security tools for autonomous AI agents

Securing autonomous AI agents

Red Hat and NVIDIA are expanding their collaboration to help businesses deploy agentic AI, which refers to autonomous systems that can reason and act independently. As organizations move toward these workflows, they require a way to ensure these agents operate safely within complex IT environments. The companies are addressing this by focusing on security-enhanced runtimes and robust hardware acceleration.

A core part of this effort is the OpenShell Reference Architecture. This initiative leverages the OpenShell project, which is an open-source sandbox runtime designed to provide standardized guardrails for AI agents. By using a sandbox, companies can isolate the agent's actions to prevent it from accessing unauthorized data or performing unintended tasks.

The partnership aims to make security a default practice rather than an afterthought. Red Hat is working to integrate the OpenShell architecture into its native platforms. This allows IT teams to maintain operational visibility and runtime safety as they scale their AI operations across hybrid clouds.

Optimizing model performance

Running autonomous agents creates significant demand for inference, which is the process of the AI generating an output from an input. To manage this, Red Hat and NVIDIA developed a joint reference design that includes Model-as-a-Service (MaaS) access controls. This system allows for self-service model discovery, fine-grained access policies, and subscription-based entitlement tracking.

The architecture also incorporates an AI Gateway and high-performance inference tools like vLLM and llm-d. These tools help organizations decouple how they govern AI access from how they actually execute the models. This separation ensures that high throughput and low latency are maintained even while enforcing strict quota management and cost attribution.

The solution specifically supports the latest NVIDIA Nemotron models, such as the Nemotron 3.5 Super VL. To make these models faster, the companies use quantization techniques through Red Hat’s LLM compressor. This process reduces inference latency while retaining accuracy, providing a transparent way for teams to generate optimized checkpoints for their specific needs.

Hardware and CPU support

To keep up with the heavy demands of agentic AI, Red Hat is aligning its software with upcoming hardware from NVIDIA. This includes providing Day 0 enablement, which means the software will be ready to work the moment the new hardware is released. This support covers the open hybrid cloud portfolio, including the Red Hat AI Factory.

A key component of this hardware strategy is the NVIDIA Vera Rubin architecture. Specifically, Red Hat Enterprise Linux for NVIDIA now provides support for NVIDIA Vera CPUs. These processors are designed to handle the continuous reasoning loops and tool orchestration that define agentic AI workflows.

NVIDIA Vera CPUs are intended to deliver 2X faster agentic CPU performance compared to traditional processors. This improvement is meant to benefit longer-thinking agents and faster-learning RL (reinforcement learning) systems. By providing higher memory bandwidth and faster sandbox performance, the hardware addresses specific bottlenecks in data-driven AI factories.

Enterprise deployment and scale

The joint advancements from Red Hat and NVIDIA are designed to give IT decision-makers a trusted foundation for scaling AI. Because agents require a secure environment to act effectively, the partnership focuses on providing a blueprint that combines a secure runtime with governed inference. This allows companies to move from experimental AI to production-ready systems.

The tools and reference designs are available through a broad ecosystem of partners, including Cisco, Dell Technologies, and Lenovo. By working with these established hardware and software providers, Red Hat and NVIDIA ensure that their security-forward architecture can be deployed across various enterprise environments. This collaborative approach helps standardize how organizations handle agent security.

John Fanelli, vice president of Enterprise Software at NVIDIA, noted that controls belong outside the agent in the runtime and infrastructure. This philosophy guides the development of the new tools, ensuring that the boundaries of an AI agent are enforced by the underlying system rather than the agent itself. This approach provides a high-performance blueprint for the next evolution of enterprise computing.