Red Hat previews AI 3.6 around an AI Gateway and secure agent execution

Red Hat has published the first concrete preview of Red Hat AI 3.6, using a November 4 virtual event page to name an AI Gateway, secure agent execution and expanded safety controls as parts of the coming release.

The page describes the three-hour “Red Hat AI Elevate Virtual” event as an unveiling of Red Hat AI 3.6. It says the release is intended to move AI agents from experiments into governed production workloads across hybrid-cloud environments.

What Red Hat is previewing

The clearest platform change is an AI Gateway for operating self-hosted models and external model APIs through one governed service. Red Hat says the gateway will add per-tenant isolation, quotas, token metering and observability, giving platform teams a shared control point for consumption and accountability.

For agents, the preview names built-in sandboxing, runtime guardrails, access controls and support for the Model Context Protocol tooling ecosystem. A dedicated session promises a live demonstration of turning experimental agents into auditable workloads rather than leaving isolation and policy enforcement to each application team.

The event also outlines automated safety and security evaluations before deployment, plus integrated insights and granular controls during operation. Another session pairs Red Hat and NVIDIA around the isolation, governance and oversight needed to deliver agents at organizational scale.

Why it matters

This preview suggests that Red Hat AI 3.6 will concentrate on the platform boundary around models and agents: who may use them, where they execute, how usage is measured and how risky behavior is constrained. Those are platform-engineering concerns that become more important as organizations expose shared inference capacity to multiple application teams.

The AI Gateway is especially notable because it is framed as a common service for both internally hosted models and third-party APIs. If the shipped implementation matches the preview, teams could apply tenant policy and cost controls without forcing every application to build its own model-access layer.

What to watch next

The event page is a preview, not release documentation. It does not provide installation details, supported-version matrices or a general-availability date. Platform teams should therefore treat the named capabilities as a roadmap signal until Red Hat publishes product documentation and release notes.

The November 4 sessions should answer the operational questions left open: which Red Hat AI components supply the gateway and sandbox, how upgrades from 3.5 work, what telemetry is exposed for quotas and token accounting, and which controls are generally available rather than preview features.