Repository navigation
2.1.1070
🚀 Arthur Engine Release
October 27, 2026
This release expands cloud discovery across AWS, GCP, and new Vercel and Elasticsearch integrations, brings Windows AI agent detection, modernizes the trace viewer on shared components, and strengthens observability and security posture.
Discovery
Cloud Connectors
- Added the Vercel discovery connector (
vercel_ai) to automatically detect AI/LLM projects within Vercel teams by scanning project and team variables, AI integrations, Agent Runs, and observability data (#2433) - Added aws_containers, a cloud-class discovery source that inventories ECS (Fargate and EC2) and EKS container workloads across account regions, with role-based identity matching and image-based filtering (#2500)
- Enhanced the Cloud Run and GKE connector to select workloads by their Vertex AI callers, report refused API scopes as
DeniedScopeentries, and correctly report operator-managed workloads such as KServe and Ray (#2498) - Fixed Bedrock discovery to exclude roles whose model calls were all refused by IAM while still counting refused calls in scan results (#2499)
These updates broaden workload discovery across AWS and GCP environments, add Vercel support, and improve how refused and operator-managed workloads are handled.
Standalone Discovery
- Added a standalone-only Elasticsearch destination for discovered agents and scan outcomes, with
_bulkbatching, API key and basic authentication, and startup index validation (#2528) - Hardened the shared
HttpSink._check_acceptedmethod to confirm 2xx responses actually accepted documents before counting a batch as delivered, preventing silent data loss (#2528) - Linked the Jamf Pro runbook from the standalone
jamf_prosource and clarified lookback filter behavior on each Mac'sgeneral.reportDate(#2537)
Standalone discovery now supports Elasticsearch and Elastic Security as a SIEM destination with clearer Jamf Pro guidance.
AI Detection
- Adopted ai-detection-spec v0.16.2 (spec 0.2) across the endpoint and ML Engine, adding Windows support with 119 identifiers, Windows-specific path routes and registry uninstall keys, and improved catalog validation (#2532)
- Catalog configuration errors are now reported as
not_configuredinstead ofprovider_errorfor clearer troubleshooting (#2532) - Added a vendoring script (
vendor_ai_detection_spec.py) to the ML Engine that validates checksums, file integrity, and version consistency when updating from upstream (#2536)
AI agent detection now extends to Windows systems, and maintainers can keep the endpoint and ML Engine synchronized on the same spec release with automated validation.
Trace Viewer
- Refactored the trace viewer to render from shared components, consolidating trace tables, span details, and code highlighting while retaining currency, timezone, and 24-hour display settings (#2527)
- Added continuous-evaluation name search with pagination and updated annotation and feedback workflows in the trace viewer (#2527)
- Fixed a 500 error when filtering traces by a single span type combined with other span filters such as span name, span ID, or status code (#2533)
The trace viewer gains improved filtering and a more consistent, modular architecture while resolving a key filtering error.
UI Architecture
- Replaced locally maintained interface components, including SearchBar, CopyableChip, and Drawer, with imports from the shared components library at version 4.5.0 (#2534)
Component ownership is now consolidated in the shared library, ensuring consistency across the application with no change to existing workflows.
ML Engine Platform
- Published the ML Engine Docker image for linux/arm64 alongside linux/amd64, enabling deployment on AWS Graviton nodes and Apple Silicon without emulation (#2526)
- Standardized on ODBC Driver 18, removing Driver 17 and the Oracle Instant Client, with automatic fallback and logging for connectors configured for Driver 17 (#2526)
- Added optional CloudFormation parameters (
MLEngineAwsDiscoveryAllowDefaultCredentialsandMLEngineTaskRoleAssumableRoleArns) so AWS discovery sources can use the ML Engine task role without hand-edited task definitions (#2529)
The ML Engine now runs natively on arm64 infrastructure, simplifies its database driver stack, and enables task-role-based AWS discovery including org-wide hub-role scans.
Observability & Consumption
- The Arthur Platform now tracks ingest volume in bytes across GenAI and ML data sources to power a new read-only Consumption dashboard, with OTLP body size recorded per span and ML dataset aggregations computed during scheduled metrics runs (#2470)
- Updated the New Relic Python agent dependency from v13.6.1 to v14.0.0 (#2519)
Organizations can now view ingest volume metrics split by source, with the latest New Relic monitoring capabilities.
Infrastructure & Security
- Enforced a FIPS-compliant TLS policy (
ELBSecurityPolicy-TLS13-1-2-FIPS-2023-04) on the genai-engine ALB HTTPS listener to address SOC2 requirements, with TLS 1.2 and 1.3 still fully supported (#2467) - Added four RDS CloudWatch alarms to the genai-engine alarms stack for CPU utilization, freeable memory, disk queue depth, and volume bytes used, with a skip condition for bring-your-own-DB deployments (#2489)
- Added a CloudWatch Errors alarm for the GPU task-scaling Lambda to the engine alarms stack, scoped to the GPU deployment and wired to the existing alerting SNS topic (#2490)
Deployments gain FIPS-compliant TLS enforcement for SOC2 compliance along with expanded database and GPU scaling alerting for faster incident response.
Release notes generated by Louisa