AWS AI Updates: July 10, 2026

1. AWS MCP Server Adds OAuth 2.0 Authentication for AI Agents

AWS. The AWS MCP Server, which lets AI agents connect directly to AWS services and resources, now supports OAuth 2.0 authentication, letting agents connect using the industry-standard protocol without additional authentication software. The flow reuses existing AWS identities and IAM permissions and offers both interactive browser-based authorization and programmatic headless authentication, with administrative controls for token management, dynamic client registration, and CloudTrail audit visibility. This removes the need for custom authentication layers when deploying agents while keeping existing IAM policies and governance in place. Source

2. SageMaker HyperPod Adds Deep Health Checks for Continuously Provisioned Slurm Clusters

Amazon SageMaker. HyperPod now runs deep health checks on Slurm-orchestrated clusters created with continuous provisioning, validating GPU accelerator health and running hardware stress and connectivity tests on instances before workloads are assigned. As new nodes come online asynchronously, the service isolates them from job scheduling during checks, returns passing instances to service automatically, and integrates with automatic node recovery to reboot or replace failing hardware. AWS says the checks head off cases where a single unhealthy node wastes hours of compute and delays training jobs. The feature is available in all regions where SageMaker HyperPod runs. Source

3. SageMaker Feature Store Adds Batch Writes and Record Listing

Amazon SageMaker. Feature Store now offers a BatchWriteRecord API that writes multiple records across multiple feature groups in a single request, targeting online stores, offline stores, or both, with graceful handling of individual record failures and flexible time-to-live configuration. A new ListRecords API lets users page through feature group contents without knowing record identifiers in advance, supporting auditing, recovery, and lifecycle management, and offline stores can now use custom naming for Glue and Iceberg tables. AWS positions the additions as removing the need for custom tooling around large-scale feature ingestion. The capabilities are available in all regions where SageMaker Feature Store operates. Source

4. SageMaker Unified Studio Adds Custom Asset Types in IAM-Based Domains

Amazon SageMaker. Unified Studio now lets administrators in IAM-based domains define custom asset types to catalog diverse data formats such as medical imaging files, PowerBI dashboards, and PDF reports. Administrators set names, descriptions, and optional metadata forms, and assets can be enriched with glossary terms and documentation before publication, after which anyone in the domain can find them by name, type, or glossary term and request access through the same governed subscription workflow. The feature is available across all regions that support SageMaker Unified Studio. Source

5. SageMaker Unified Studio Workflows Gains 19 Operators for Bedrock and S3 Vectors

Amazon SageMaker. Unified Studio Workflows added 19 new Apache Airflow operators for Amazon Bedrock, S3 Tables, S3 Vectors, AWS Glue Data Catalog, and MWAA Serverless, letting builders add tasks through the visual workflow creator without writing custom integration code. The operators can manage Bedrock guardrails, provision S3 resources, handle Glue Data Catalog operations, and run MWAA Serverless workflows, reducing the need to switch between consoles or author custom DAG code for multi-service orchestration. The feature is available in all regions that support SageMaker Unified Studio. Source

6. AWS Neuron 2.31.0 Ships NKI 0.5.0 and UltraServer Operator Beta

AWS. Neuron 2.31.0, the SDK for AWS Trainium and Inferentia accelerators, adds NKI 0.5.0 with MX FP8 scale dtype support, tensor indirection on compute operations, and tensor view APIs for layout transformations, alongside a redesigned compiler backend that improves performance on Trn2 and Trn3 instances. The release includes 14 new experimental kernels targeting MoE training, deformable attention, and ring attention, and it introduces the Neuron UltraServer Operator for Amazon EKS in public beta to automate server discovery and workload distribution for Trainium deployments. The software is available in all regions where EC2 Trn1, Trn2, Trn3, Inf2, and Inf1 instances run. Source