Red Hat AI 3.5 adds safety tools for enterprise AI

Red Hat has released Red Hat AI 3.5, adding new safety and observability tools for AI deployments.

The update targets organisations moving AI systems from pilot projects into broader production use across hybrid cloud environments. It includes model evaluation, usage monitoring, controls for shared GPU infrastructure, and tools for building and managing AI agents.

One of the main additions is the general availability of EvalHub, which can automate benchmarking and compliance reporting for custom models, retrieval-augmented generation systems, and agents. Evaluated models in Red Hat's catalogue now also include Garak benchmark scores, along with indicators for risks such as toxicity and exposure of personally identifiable information.

The release also expands observability. New dashboards show inference health, GPU use, and model performance in real time, while non-administrator users can view token consumption on a per-user basis. It also adds visual tracing for agent-based workflows through MLflow.

Shared resources

Another focus is the management of scarce computing resources, particularly GPUs used for AI inference. Red Hat AI 3.5 introduces fair-share scheduling to divide GPU resources across tenants and priority-aware serving to route requests by importance. The aim is to protect time-sensitive workloads while allowing background jobs to use spare capacity.

For customers that need stronger separation between tenants, the software now officially supports hosted control planes on OpenShift Virtualization. In that setup, each tenant gets a dedicated cluster control plane while sharing the underlying hardware. AI workloads can also run in virtual machines on shared GPU-enabled infrastructure.

These options are intended for internal platform teams and service providers that want to offer AI as a shared service while retaining central oversight of upgrades and infrastructure operations.

Agent tools

Red Hat AI 3.5 also expands features for AI agents, an area of growing interest for companies seeking to automate tasks such as document handling, research, and software review. The Responses API is now generally available, providing a standard interface for multi-turn agent interactions.

The release includes built-in NeMo Guardrails, which can intercept malicious tool calls, and AutoRAG, which links enterprise data repositories to agent applications. AutoRAG adds multilingual document support, conversational testing, and contextual retrieval.

Inference-Time Scaling is another new element. It adjusts compute use dynamically based on the complexity of a query, which could help organisations manage AI infrastructure costs more closely when running reasoning-heavy workloads.

Pre-built agent templates have also been added in AI Hub. These cover code review, document processing, and research workflows, giving teams reference implementations and deployment configurations for common enterprise use cases.

Model catalogue

Red Hat has added more than 20 validated models to its catalogue, including models from Google, NVIDIA, and Alibaba Cloud. Some are tagged as validated for tool-calling, which is relevant for agent-based applications where models need to interact with software tools and external systems.

Other additions include controlled model rollout for managing traffic during model updates, the general availability of CPU offloading, and a developer preview of storage offloading for handling longer conversations and larger documents without adding more GPU hardware.

Red Hat is also extending distributed inference beyond OpenShift to third-party Kubernetes services. Support is now generally available on CoreWeave CKS and Microsoft Azure, while Amazon EKS remains in technology preview.

A separate developer preview brings the Kubeflow Spark Operator into the workbench environment, allowing distributed data processing alongside model serving on the same platform.

The release reflects a broader market shift as companies move from testing generative AI models to putting them under the same governance and operational controls as other production systems. That has increased demand for tools that measure model behaviour before deployment, monitor usage after launch, and allocate expensive hardware across multiple teams.

Joe Fernandes, Vice President and General Manager, AI Business Unit, Red Hat, described that shift in the company's launch remarks.

"The conversation has moved from getting AI into production to running it at scale as trusted enterprise infrastructure, which requires safety evidence, governed agents, cost attribution and multi-tenancy," said Joe Fernandes, Vice President and General Manager, AI Business Unit, Red Hat. "With Red Hat AI 3.5, we are delivering the operational controls, verifiable trust and agentic foundations IT leaders need to run AI as a safe, controlled and accountable enterprise AI architecture across the hybrid cloud."