Red Hat has published updated images, but the linked 3.4 release notes still describe 3.4.4 and the advisory says upgrade documentation will be refreshed shortly.
An open pull request reports as much as 1.92× faster non-speculative decode on GB300 while narrowing where agent-assisted kernel code can run.
CVE-2026-99999 is rated Important by Red Hat, but the vendor’s active-exploit designation and lack of a listed fix make it an immediate OpenShift watch item.
The implementation supports tool calling and structured output, while vLLM reports a 2.71× batch-one decode speedup from seven model-specific changes.
Connectivity Link 1.5 adds tenant isolation, protocol negotiation, progressive tool discovery, audit telemetry and NeMo guardrails for agent tool traffic.
A merged model-integration patch and six new recipes put typed multimodal decisions behind vLLM’s System One API, although users need a nightly build for now.
An online-fitted estimator lets more speculative-decoding setups trim low-value draft tokens without retraining existing checkpoints.
The Kubernetes deployment adds API-key authentication, subscription-aware access and token quotas in front of vLLM and KServe model endpoints.
Ferrocene 26.05 supports a qualified RHIVOS 2 target now, with the full joint solution planned for the second half of 2027.
The first Quarkus 4 beta raises the Java baseline, refreshes the networking and persistence stack, and asks extension maintainers to move before a planned late-November final release.
A new OpenAI-compatible embeddings endpoint lets teams generate vectors, discover compatible models and keep retrieval and chat behind one managed project boundary.
Red Hat’s lifecycle record puts the 3.3 release past its full-support end date; customers should verify a supported channel and upgrade path rather than infer an undocumented next phase.
Apache Camel’s latest benchmark round adds discovery and validation support for Kamelets, lifting a local model’s pass rate from one task in 15 to 47 in 50.
A single Podman image packages offline recommendations, drift checks and must-gather analysis for restricted environments.
The Satellite-integrated service keeps system data local while combining fleet risk views, CVE prioritization and remediation paths.
CVE-2026-105192 has no published fix; operators using LMCache’s multiprocess mode should keep its ZeroMQ port off routable networks.
Operators should update current vLLM and Starlette deployments after an audit linked vLLM’s most severe finding to a vulnerability now listed as exploited in the wild.
Red Hat has packaged multicluster management, security, registry and storage tools as consumption-based marketplace offers for its managed OpenShift services.
IBM and Red Hat are opening a request-driven path for enterprises to obtain reviewed, backported fixes for older open source dependencies.
A distributed AxyomCore deployment pairs a central core in Fornebu with an autonomous edge core in Svalbard, using OpenShift as the common platform layer.
The October lifecycle transition narrows the kinds of fixes Red Hat will accept and gives platform teams a dated migration boundary.
The customer-managed release can keep code and inference inside approved environments, but buyers must supply the OpenShift capacity and a supported model.
The Technology Preview adds declarative sandbox lifecycles, warm pools and optional VM-level isolation, but keeps production support off the table.
Red Hat plans to make RHCOS 10 the default for new OpenShift installations in Q4, requiring early checks for privileged containers, UBI7 workloads and older x86 hardware.
The policy gives each GA build eight months of full support, four months of maintenance, and a version date that operators can use to calculate end of life.
The company says it will contribute Linux, Kubernetes, security and distributed-systems expertise while testing the project against its own internal agent deployments.
The storage vendor says OpenShift Virtualization customers can now use synchronous or asynchronous replication, failover and S3 backups without a separate DR product.
OpenShell and DOCA are coming to Red Hat AI Factory, while NVIDIA’s BlueField-based Sentry design adds an independent enforcement plane for agent workloads.
The extension accepts the July 2026 stateless protocol and legacy session-based clients on one endpoint, giving platform teams a staged migration path.
The enterprise installation runs as a tenant workload on customer-managed OpenShift clusters, with customers supplying the model endpoints and operational controls.
SCAP Security Guide 0.1.82 aligns the RHEL 10 profile with the official benchmark and makes it usable across OpenSCAP, Satellite and Red Hat Lightspeed.
The new Gumbel-max implementation embeds a detectable provenance signal during generation while preserving sampling behavior and speculative-decoding performance.
The certification opens the managed OpenShift service to higher-sensitivity CUI and unclassified national-security workloads in Microsoft Azure Government.
The integration targets teams moving virtual machines onto OpenShift Virtualization without creating a separate storage and operations silo.
The designation brings regular testing and bug-fix priority to RHEL, giving robotics teams a supported enterprise Linux target for ROS 2 development and deployment.
The open-source Lola module runs in Claude Code, Cursor and Gemini CLI, tying proposed attack scenarios to files and lines while keeping human review in the loop.
New RHEL subscription features add CrowdStrike YARA coverage, SIEM and SOAR streaming, Event-Driven Ansible hooks, and visual SCAP policy migration.
The September patch maps Git command-injection and Kubernetes token-exfiltration fixes to specific Ansible Automation Platform components and operator builds.
A verified Red Hat solution says a migrated guest interface can gain an “_bkp” suffix and lose connectivity, making post-cutover network checks essential.
The proof of concept benchmarks embedding models, promotes the winner and deploys a CPU-based semantic-search service through KServe.
The early architecture links policy extraction, risk mapping, red-team scenarios and Kubernetes-ready mitigations, but much of the workflow remains a roadmap.
The new PyNvVideoCodec backend shifts frame decoding away from host CPUs; vLLM reports more than twice the throughput at eight H100 GPUs in its captioning test.
The project now validates route examples during the build and exposes documentation as Markdown, while its redesigned front page offers runnable starts for Camel CLI, Spring Boot and Quarkus.
A new five-billion-parameter draft model and hidden-state connector turn speculative decoding into a capacity-planning choice across separate inference and training nodes.
Chord adds shape-aware W4A16 CUDA paths for Kimi K2.x, while its grouped vLLM integration remains unfinished.
A Red Hat case study says the migration produced a working prototype in two weeks while adding a controlled path between informational and consultative modes.
Red Hat says its membership service supplies tested, digitally signed remediations and rebuilt libraries for exact dependency versions, including Java and Python.
The HyperShift-based option removes control-plane compute from customer subscriptions, separates control-plane and worker upgrades, and starts with CLI deployment.
A Red Hat engineering demo uses Backstage catalog data, TechDocs and software templates to ground an AI coding agent before it proposes architecture or code.
Red Hat says nodes may move to cgroups v2 during an upgrade from 4.18.46 even when the cluster configuration explicitly requests v1.