August 2026 Release Notes

Arthur Engine & Toolkit

Source trace drawer. A new side-panel drawer displays linked traces directly within dataset and experiment views, allowing trace investigation without navigating away.
Continuous Eval Name filter. The filter is now a searchable dropdown populated from the task's continuous evaluations, while still supporting free-text substring search.
Multi-select trace filters. Continuous Eval Run Status and Status Code filters now accept multiple selections, returning any trace matching at least one selected value.

Deep-linking to dataset rows. Row IDs in the experiment "Dataset Row Data" modal are now clickable links that open the specific row in a shareable, browser-back-compatible URL.
Model whitelist for providers. Admins can curate which models appear in provider dropdowns via new GET/PUT endpoints and a Visible models section under Settings → Model Providers, with full backward compatibility.

GPU-metric autoscaling add-on chart. A new standalone Helm chart for GenAI Engine on EKS Auto Mode + Karpenter drives pod autoscaling from GPU utilization and memory metrics via a DCGM → Prometheus → prometheus-adapter → HPA pipeline.
External HPA management flag. A backward-compatible flag in the core genai-engine chart prevents conflicts when HPA is managed outside of Helm.
CloudFormation worker count default. GenaiEngineServerWorkerCount is now defaulted to 3 workers across all GPU CloudFormation templates, delivering approximately 44% more throughput on g4dn.2xlarge instances with no additional configuration.

Gunicorn upgraded to v26. Adds HTTP/1.1 request validation, request-smuggling hardening per RFC 9110/9112, and expanded ASGI compatibility; the eventlet worker is removed and requires migration to gevent, gthread, or tornado.
Node.js base image updated to v24-alpine. Delivers performance improvements, new crypto algorithms, and enhanced HTTP capabilities.
SSL verification variable renamed. KEYCLOAK_SSL_VERIFY has been renamed to more accurately reflect its role, with full backward compatibility for existing configurations.
OpenAI client updated to v3. Migrates to HTTPX2 as the default HTTP client; applications using custom HTTPX configurations must update to remain compatible.
gcsfs updated to v2026.8.0. Introduces adaptive concurrent prefetching for dramatically improved sequential read throughput on ML/AI workloads.
fsspec filesystem packages updated to 2026 versions. Now default to Azure credentials and remove ADLS Gen1 support in favor of the az:// protocol.

cryptography updated to v50.0.0. Addresses three CVEs in X.509 verification and PKCS#7 handling, reducing open Trivy findings from 11 to 3.
Build-time Python tooling removed from ml-engine runtime image. Drops pip, setuptools, and uv from the runtime image, eliminating two CVEs and closing all remaining Trivy findings.
Dependabot alerts resolved. Resolved 9 of 10 open alerts across four projects and restored dependency resolution in genai-engine.
pypdf updated to v6.15.0. Fixes two medium-severity DoS vulnerabilities in font processing and text extraction.
3-day dependency cooldown enforced at Yarn install layer. Extends release-age soak protection to yarn add and transitive resolution while exempting first-party @arthur/* packages.
Renovate auto-merge resumed. Auto-merge and auto-approve are restored for low-risk updates, retaining the 3-day soak, required CI checks, and human review for major bumps.
Renovate auto-merge paused during supply-chain campaign. As a precaution during the active keyv/Shai-Hulud npm supply-chain campaign, non-major updates were held for manual review.

framer-motion replaced with motion. The recommended successor package from the same maintainer is now used, preserving full animation compatibility.
motion updated to v12.43.0. Adds hardware acceleration for backgroundColor and SVG elements and fixes an AnimatePresence regression that caused exiting children to remount when new children were added.

Fixed the Renovate auto-fix workflow git push authentication by explicitly passing the Automator App token, unblocking automated code fixes.
Fixed an auto-fix approval race condition by pinning approvals to the exact post-push commit SHA.
Fixed the observability-SDK build by moving the README into the package directory for compatibility with hatchling 1.32.
Stabilized CI against uv network flakes when fetching the en-core-web-lg wheel by adding retries, capping concurrency, and refining trigger filters.
Fixed LlmChatCompletion tracking for Gemini with LangChain tools in the New Relic agent (v13).
Fixed incorrect ls() results for shared-prefix files in adlfs (v2026.8.0).