Three Red Hat engineering posts define complementary tests for guardrail policy, tool-call execution and failures that only emerge across complete agent traces.
A Red Hat account of its own IT cleanup puts dashboard retirement, spreadsheet consolidation and OpenShift standardization ahead of model deployment.
Red Hat’s promotion removes year-one subscription charges, but customers still carry migration costs and must obtain private eligibility and later-year pricing before comparing offers.
A four-step FastH3 profile produced complete 10-second audio-video files in about 8.7 seconds on eight NVIDIA B300 GPUs, but the project says its raw benchmark bundle is not yet public.
The proposed deployment record combines adversarial red-team results, PII-exposure scores and toxicity evaluations instead of treating benchmark accuracy as a safety certificate.
A reproduced attack turns a small adapter into a covert exfiltration mechanism, shifting the defensive focus from weight inspection to deployment, network and provenance controls.
Red Hat’s support assistant uses live product content, Granite Guardian and independently measured skill routing to keep troubleshooting answers inspectable.
Red Hat’s ART pipeline generates tailored adversarial prompts, escalates attacks only when cheaper probes fail, and tracks the results through OpenShift AI.
Three recent artifacts show platform teams making separate placement decisions for accelerators, batch work and media-aware requests rather than treating every inference call alike.
The code is on vLLM’s main branch and in a model-specific image, but it arrived after v0.28.0 and is not yet part of a newer general release.
Red Hat’s upstream prototype verifies Sigstore bundles during Kubernetes reconciliation, complementing SPIFFE runtime identity without yet blocking unsigned legacy cards.
The OpenShift-based design separates perception and planning from execution, then routes approved work through distinct digital and physical orchestrators.
Three new engineering guides show how managed APIs, self-hosting and shared inference move responsibility across the same four-layer AI stack.
The bank’s new stack shifts deployment and administration toward a consistent container platform, but the published performance gains remain vendor-reported case-study results.
The Emerging Technologies team points agent builders toward structural instruction/data separation and privilege boundaries, while warning that no current technique is foolproof.
A Red Hat engineering post replaces one oversized prompt with narrow agents, board state and isolated Git worktrees—and documents the operational gaps that remain.
AT&T used Red Hat’s SDG Hub to turn standards documents into a 440-billion-token training set, then ran full-weight fine-tuning on AMD infrastructure.
The release moves referenced-task reads behind a fail-closed boundary, adds application-controlled stream cleanup and requires import and task-state migration work.
The Prague agenda points application and platform teams to real-time risk events, cross-domain identity and MCP-era authorization—but it is a conference programme, not a release roadmap.
A 90-repository review found 572 cryptographic findings, but the first fixes focus on removing hardcoded algorithms and TLS barriers rather than replacing the stack.
The framework asks teams to measure traces, memory and state changes—not just outputs—and to isolate every trial before trusting the score.
SemiAnalysis has released an Apache 2.0 benchmark and trace-replay dataset aimed at multi-turn coding agents, with vLLM and llm-d contributors involved in the optimization work.
A developer-preview capability combines model-based validation, optional human approval and systemd isolation, but Red Hat still draws the line at local troubleshooting.
A Red Hat field build compares strict, standard and permissive agent policies—and makes their remaining permissions visible.
Dataverse Agent combines governed data products, staged SQL generation and MCP interfaces on OpenShift, with three reusable templates now available.
The Red Hat prototype compares agent-card claims with synthetic probes and MLflow traces, but remains a post-hoc audit rather than runtime enforcement.
The new BGP EVPN support gives primary cluster user-defined networks a standards-based path into existing data-center fabrics, with virtualization migrations as the clearest early use case.
The maintainers’ six-to-12-month plan targets delegated identity, asynchronous work, a common transport model and more consistent tool results.
AgentTrust, the DCI MCP server and Code-to-Docs show three places platform teams can turn an AI workflow’s promises into enforceable controls.
The DCI MCP server offers a broad root-cause evidence path, while making transport, token scope and write controls part of the deployment design.
A joint RPI–IBM project published by Red Hat Research argues that controllable forecasting matters more than marginal leaderboard gains.
The API surface is marked generally available, but the 3.5 EA2 deployment carrying it has no production support or upgrade path; MaaS and evaluation additions also need separate preview treatment.
The Red Hat AI Americas field build combines MCP access, scoped storage, provenance and audit controls in a reference architecture for teams running many agents.
The financial network’s OpenShift security design replaces cluster polling with stateless aggregation and sandbox regression tests.
The upstream prototype combines build-time attestations with SPIFFE workload identity, while leaving unsigned legacy cards in a migration path.
The project replaced a four-signal default with bottleneck-matched routing that operators can calibrate for each model, accelerator and engine.
Redis routing, workflow navigator sizing and realized token accounting form one loop: reduce avoidable demand, fit capacity to the remaining work, then verify the economics in production.
The developer preview ranks models and emits KServe configuration from stated constraints; it does not replace workload-specific benchmarking or deployment review.
OpenShift 4.22 previews RHCOS 10, while a future OSStreams API is intended to let RHCOS 9 and 10 nodes coexist during migration.
The Red Hat AI Americas reference build makes MaaS evaluation and access controls inspectable, while its defaults stop short of supporting broad claims about abliteration.
Red Hat’s new explainer separates single-node vLLM efficiency from the distributed routing and orchestration job assigned to llm-d.
Distributed Layerwise Offload cuts device and host-memory pressure, but its working quickstart requires versions newer than llm-d’s current bundle.
Across policy enforcement, certificate rotation and a RHEL VM factory, automation acts inside inventory, approval and lifecycle boundaries that people still own.
The Dutch government platform serves 48 organizations across 11 ministries and is extending its standardized OpenShift foundation toward scarce GPU infrastructure and governed AI development.
Red Hat’s new open source project maps policy to risks, test scenarios and deployable mitigations, aiming to make agent approval repeatable rather than manual.
The system couples MCP-governed tools, hybrid model routing and a 300,000-document retrieval pipeline with ordinary platform controls.
A thin-image model separates Day 0 builds, Day 1 hardening and Day 2 lifecycle control across GitLab, Ansible and Satellite.
Lightspeed’s GA support boundary and expanded MCP scaffolder operations give platform teams more leverage—and more policy surface—to manage.
The insurer says standardized cluster blueprints, distributed ownership and hosted control planes helped move 1,500 workloads in 10 months.
Red Hat proposes OPA checks, evidence collection and event-driven remediation as one control loop—but AI actions are governed only when they actually pass through it.