live wire
▸AI · EAGLE-4 0.1.4 fixes multi-GPU vLLM collective ordering during startupEAGLE-4▸SECURITY: Apache Camel 4.23.0, 4.22.2 and 4.18.5 fix eight component vulnerabilitiesApache Camel▸AI · Red Hat shows DRA doubles llm-d KV-cache throughput by aligning GPUs and RDMA NICsRed Hat Developer▸Red Hat Developer Hub 1.9.10 ships as a Critical security updateRed Hat errata▸OpenShift Dev Spaces 3.31 adds live editor switching and DevWorkspace Operator 1.0Red Hat documentation▸OpenShift Pipelines 1.24.1 tightens Manual Approval Gate RBAC and approval integrityRed Hat release notes▸AI · Red Hat survey: 97% of UK CIOs trust agent security, but 43% can trace data flowsRed Hat▸JAVA · IBM and Kaggle launch a living benchmark for agentic Java framework migrationsIBM Research▸AI · Red Hat tests policy-derived data controls for AI agentsRed Hat▸Apache Camel pattern turns routes into bounded LLM tools and MCP servicesRed Hat Developer▸Atos tunes small models on OpenShift AI for structured German-language analysisRed Hat Developer▸Red Hat publishes checklist for evaluating third-party OpenShift applicationsRed Hat Developer▸SECURITY · OpenShift 4.21.40 patches OpenSSL, OpenSSH and PCRE2 flawsRed Hat Product Security▸SECURITY · Apache Camel fixes A2A webhook SSRF gap in 4.22.2 and 4.23.0 (Oct. 23)Apache Camel▸AI · EAGLE-4 0.1.4 fixes multi-GPU vLLM collective ordering during startupEAGLE-4▸SECURITY: Apache Camel 4.23.0, 4.22.2 and 4.18.5 fix eight component vulnerabilitiesApache Camel▸AI · Red Hat shows DRA doubles llm-d KV-cache throughput by aligning GPUs and RDMA NICsRed Hat Developer▸Red Hat Developer Hub 1.9.10 ships as a Critical security updateRed Hat errata▸OpenShift Dev Spaces 3.31 adds live editor switching and DevWorkspace Operator 1.0Red Hat documentation▸OpenShift Pipelines 1.24.1 tightens Manual Approval Gate RBAC and approval integrityRed Hat release notes▸AI · Red Hat survey: 97% of UK CIOs trust agent security, but 43% can trace data flowsRed Hat▸JAVA · IBM and Kaggle launch a living benchmark for agentic Java framework migrationsIBM Research▸AI · Red Hat tests policy-derived data controls for AI agentsRed Hat▸Apache Camel pattern turns routes into bounded LLM tools and MCP servicesRed Hat Developer▸Atos tunes small models on OpenShift AI for structured German-language analysisRed Hat Developer▸Red Hat publishes checklist for evaluating third-party OpenShift applicationsRed Hat Developer▸SECURITY · OpenShift 4.21.40 patches OpenSSL, OpenSSH and PCRE2 flawsRed Hat Product Security▸SECURITY · Apache Camel fixes A2A webhook SSRF gap in 4.22.2 and 4.23.0 (Oct. 23)Apache Camel
upstreambeat.ai
releaseAI

OpenShift AI 3.5 makes upgrades explicit and evaluation production-ready

The release adds operator-enforced upgrade checks, a supported EvalHub control plane and production controls for distributed inference.

By The News Desk· Oct 25, 2026the quick take — two AI hosts go live when you do

Red Hat’s release notes now identify OpenShift AI 3.5 as generally available. The release is broad, but its clearest theme is operational discipline: administrators get explicit upgrade gates, evaluation moves into a supported production service, and distributed inference gains controls for mixed workloads and safer rollouts. The 3.5 release notes document all three changes.

Upgrades now stop for required preparation

During an upgrade to 3.5, the OpenShift AI Operator runs checks before updating platform components. Gates that require review, migration or cleanup are recorded in the odh-upgrade-acks ConfigMap, and the operator acknowledges resolved gates during reconciliation. Red Hat also added a migration guide for moving from OpenShift AI 2.25.9 and later 2.25 releases to 3.5 with rhai-cli, covering both side-by-side and in-place approaches and component-specific migration actions. After the upgrade, administrators must move away from the temporary support-required-upgrade-3.5 channel to an appropriate supported 3.5 channel. Red Hat describes the checks and channel step in the release notes.

EvalHub becomes the supported evaluation layer

EvalHub is GA in 3.5, with a documented API versioning and breaking-change policy, Red Hat support and must-gather diagnostics. The release includes a Python SDK and CLI for submitting jobs, discovering providers and retrieving results, plus an adapter SDK for custom evaluation frameworks. Namespace administrators can also deploy tenant-specific EvalHub instances, while the operator can create a ServiceMonitor so platform Prometheus scrapes EvalHub metrics automatically. Those capabilities are listed as 3.5 GA features.

The same evaluation stack now underpins automated adversarial scans for newly validated models, with quantitative vulnerability results published alongside catalog entries. Automated Red Teaming is also GA, including disconnected operation and execution through EvalHub or Kubeflow Pipelines. Red Hat says the scans use Garak and EvalHub.

Distributed inference gets production traffic controls

For llm-d deployments, priority-based flow control is GA. Operators can define priority tiers, saturation detection and queuing policies so latency-sensitive traffic can take precedence over batch work. Deployments upgrading from the 3.4 Technology Preview must account for API and metrics-prefix changes. OpenShift AI 3.5 also adds weighted controlled deployments for comparing inference workload versions, inference-aware pod lifecycle handling, multimodal cache-aware routing and OpenTelemetry-compatible tracing. KServe RawDeployment mode separately gains canary rollout support through weighted OpenShift Route or Gateway API backends. The serving changes and upgrade cautions are detailed in the release notes.

What administrators should check

Before scheduling the upgrade, teams should run the migration assessment and inventory deprecated or removed components. The release deprecates FMS Guardrails Orchestrator in favor of NeMo Guardrails and the standalone LM-Eval service in favor of EvalHub. It removes RStudio Server workbench images, Kubeflow Training Operator v1 runtime images, and tf2onnx from TensorFlow images. Red Hat’s support-removals section explains each migration or removal.

sources

Filed by The News Desk. Corrections: desk@upstreambeat.ai · Our standards →

comments · 0

    Comments are moderated before they appear. Your email is used once to confirm it is you — never shown, never sold. Corrections and questions get an answer from the desk when we have one.

    more from the desk