AWS AI Updates: August 21, 2026

1. SageMaker AI Studio Adds Generative AI Inference Recommendation

AWS. Amazon SageMaker AI Studio added a visual, low-code interface that benchmarks multiple inference configurations on real GPU infrastructure using NVIDIA AIPerf and returns ranked, production-ready recommendations. Practitioners specify whether to prioritize latency, throughput, or cost, and the system applies techniques such as speculative decoding or kernel tuning automatically. AWS positions it as reducing deployment optimization work from weeks to hours by removing manual configuration guesswork. Source

2. Amazon Bedrock Web Search Adds External Web Access Control

AWS. Amazon Bedrock’s Web Search feature added an external_web_access parameter that lets models retrieve content directly from the public internet for time-sensitive grounding such as live scores and pricing. Teams handling sensitive data can set the parameter to false to keep retrieval within their AWS boundary with no data egress. This gives developers a single control to balance current-information needs against data privacy requirements. Source

3. Bedrock AgentCore Web Search Adds Domain and Date Filtering

AWS. AgentCore Web Search now lets agents pass include/exclude domain lists and publication date ranges directly in each tool call, giving per-request control over sources and time windows without administrative reconfiguration. This helps regulated industries and research workflows enforce governance over information provenance and content recency. Support also expanded to the Europe (Ireland) and Asia Pacific (Tokyo) regions. Source

4. SageMaker Notebooks Support Trusted Identity Propagation

AWS. SageMaker Notebooks added Trusted Identity Propagation for Amazon Athena, Amazon Redshift, and Amazon EMR Serverless, flowing each user’s IAM Identity Center identity through to AWS Lake Formation for per-user data access. This enforces individual row-level and column-level boundaries instead of relying on shared execution roles, and produces CloudTrail attribution showing which user accessed specific data. Identity propagates automatically through existing compute connections without separate login or token handling. Source

5. Cost Anomaly Detection Now Monitors Third-Party Bedrock Models

AWS. AWS Cost Anomaly Detection extended its machine learning monitoring to third-party foundation models on Amazon Bedrock, including Anthropic Claude, through the AWS managed service monitor with no setup required. When spending deviates unexpectedly, it sends alerts with breakdowns ranked by dollar impact across service, account, region, and usage type. This helps teams running production generative AI workloads catch runaway model costs without building manual tracking. Source

6. Amazon Bedrock Adds SpaceXAI Grok 4.6

AWS. Amazon Bedrock now offers SpaceXAI Grok 4.6, a frontier model positioned for coding, agentic, and knowledge tasks, with US Geo and Global cross-Region inference profiles. The cross-Region routing automatically distributes requests across multiple AWS Regions for higher throughput and lower per-token cost. Geography-specific profiles let teams scale without manual capacity management while keeping data residency controls. Source

7. Amazon Quick Adds Deny-by-Default Governance for AI Capabilities

AWS. Amazon Quick added a governance option that lets administrators restrict the AI capability category in a custom permissions profile so new AI features are denied by default rather than automatically available. This lets teams evaluate and explicitly approve capabilities before they reach users instead of reacting after rollout. It gives organizations tighter control over AI feature adoption for security and compliance. Source

8. AWS Security Agent Adds Budget Controls and Finding Revalidation

AWS. AWS Security Agent, now part of AWS Continuum, is an on-demand penetration testing service that uses an AI agent to autonomously test web applications for vulnerabilities. Teams can now set a maximum task-hours limit per test, with billing tracking actual task-hours consumed so a higher cap does not raise costs unless the test needs the time. A new finding revalidation capability re-tests specific vulnerabilities after fixes without rerunning a full test, streamlining remediation verification. Source