New AWS AI and Cloud Features: What to Change in Agent Security, Data Access and Operations

What Happened

AWS announced a broad set of AI, data and infrastructure changes on November 9–10, 2026. The most consequential updates for production teams are:

AI applications and agents

• Web search: On November 10, Amazon Bedrock AgentCore made its InvokeWebSearch API generally available. It returns ranked results and extracts from Amazon’s web index, supports domain and publication-time filters, and requires no search resource to provision. It is available in six Regions and costs $7 per 1,000 queries. [1]
• Models: On November 9, Claude Fable 5.2 became available through Amazon Bedrock and Claude Platform on AWS, including in AWS GovCloud (US). Claude Mythos 5.2 became available in GovCloud to organizations allow-listed through Anthropic’s Cyber Verification Program. Both are designated covered models with safety-monitoring data-retention requirements; eligible Fable customers have a temporary zero-data-retention option. [6] [7] [12]
• Agent isolation: AWS Nitro Sandboxes now place EKS pods, ECS tasks or lightweight VMs in separate hardware-isolated VMs on shared EC2 instances. AWS says the capability is available on most Nitro-based instances in commercial Regions at no additional charge. [13]
• Product and workflow changes: Amazon Quick and Amazon Quick Sight were renamed AWS Quick and AWS Quick Sight, with no customer action or pricing change. Amazon Connect Customer added customizable evaluation templates for human and AI-agent interactions. [5] [8]

Data and analytics

• Integration: AWS Glue added 100 native SaaS and REST connectors for Glue 5.0 and later. AWS also announced a bidirectional, zero-copy Workday Data Cloud integration on November 10, while S3 Access Points for FSx expanded to additional Regions so S3 applications can access file data without copying it. [3] [15] [16]
• Visibility and graphs: Athena added service-managed Iceberg system tables, including query_executions, for recent usage and audit analysis. AWS released a locally run, open-source UI that projects selected Iceberg data through Athena into Neptune Analytics graphs. [9] [11]
• Database performance: Amazon DocumentDB Serverless platform version 4 offers up to 35% higher throughput. Existing version 3 clusters can upgrade through a maintenance action, but cannot roll back. [10]

Infrastructure and governance

• Controls: CloudWatch Logs added account-level retention policies, with up to 20 per Region and the longest matching retention period taking precedence. AWS STS added cross-partition federation using short-lived web identity tokens. Route 53 Profiles now permits multiple priority-ordered Profiles per VPC. [2] [4] [14]
• Accelerators: EKS Auto Mode and Karpenter added Dynamic Resource Allocation for NVIDIA GPUs and Elastic Fabric Adapter devices. EKS Auto Mode requires Kubernetes 1.36 or newer; the support excludes GovCloud (US) and China. [17]

Why It Matters to Businesses

These releases can remove custom integration and operations work, but they do not remove the need for application controls. AgentCore web search provides a managed retrieval call, not a guarantee that returned pages are trustworthy. Glue connectors, Workday federation and FSx access can shorten data paths, but each introduces permissions and usage that must be governed across services. [1] [3] [15] [16]

For regulated teams, the model-access and isolation announcements deserve separate reviews: a model’s availability in GovCloud does not settle whether its retention terms fit a workload, and a hardware-isolated sandbox does not replace restrictions on network access or credentials. [6] [7] [13]

Kimbodo Engineering Perspective

We would prioritize changes that replace brittle custom code or close a measurable control gap. Account-level log retention and Athena system tables are relatively contained operational improvements. Switching an agent to web search, adopting a new model, or changing workload isolation has a wider effect on data exposure, cost and failure modes, so each needs a workload-specific test. [1] [2] [11] [12] [13]

Version upgrades need an explicit rollback decision. DocumentDB platform version 4 is irreversible for an upgraded cluster; its “up to 35%” throughput claim should be tested against the application’s own queries before scheduling maintenance. [10]

How We Would Implement It

• Establish a baseline: Inventory affected Regions, workloads, model-retention requirements, log groups and data owners. Apply a default CloudWatch retention policy, then longer policies for security-related prefixes where needed. [2]
• Pilot one agent path: Put InvokeWebSearch behind an application tool interface with IAM-scoped access, domain filters where appropriate, result citations, timeouts and a query budget. Log tool calls through CloudTrail data events; test how the agent handles malicious or irrelevant page text. [1]
• Prove data access before migration: Trial one Glue connector or federated Workday query with least-privilege permissions on both sides. Compare freshness, latency and query cost with the existing pipeline. Enable Athena system tables to inspect usage after rollout. [3] [11] [15]
• Validate runtime changes: Test Nitro Sandbox attestation and workload compatibility before moving untrusted agent code. For GPU workloads, pilot ResourceClaims and verify device allocation and scaling behavior before replacing device-plugin configurations. [13] [17]

Risks, Costs and Security

AgentCore search has a per-query charge; Route 53 Profile–VPC associations retain standard charges. “No additional charge” for features such as Nitro Sandboxes or CloudWatch retention policies does not eliminate underlying compute or log-storage costs. Set budgets and measure usage before broad rollout. [1] [2] [13] [14]

Review model safety-monitoring retention before sending sensitive prompts, enforce least privilege across new connectors and federation paths, and treat web results as untrusted input. For cross-partition STS access, audit token use in both partitions. For priority-ordered DNS Profiles, test which configuration wins when teams associate multiple Profiles with a VPC. [1] [4] [12] [14]

Sources

1. [1] Amazon Bedrock AgentCore now offers a Web Search API
2. [2] Amazon CloudWatch Logs announces account-level retention policies
3. [3] AWS Glue adds 100 new native connectors for popular SaaS applications
4. [4] AWS STS now supports cross-partition federation for AWS identities
5. [5] Amazon Quick is now AWS Quick
6. [6] Claude Fable 5.2 is now available on AWS GovCloud (US)
7. [7] Amazon Bedrock now offers Claude Mythos 5.2 to Eligible Customers in AWS GovCloud (US)
8. [8] Amazon Connect Customer launches evaluation form templates
9. [9] Introducing an open-source UI for graph analytics over data lakes with Amazon Neptune Analytics
10. [10] Amazon DocumentDB Serverless now provides up to 35% higher throughput
11. [11] Amazon Athena introduces system tables for observability, auditing, and usage analytics
12. [12] Claude Fable 5.2 is now available on AWS
13. [13] Introducing AWS Nitro Sandboxes on Amazon EKS, ECS and EC2 for hardware isolated pods, tasks and VMs
14. [14] Amazon Route 53 Profiles now supports multiple Profiles per VPC
15. [15] AWS Analytics services now support Workday Data Cloud integration
16. [16] Amazon S3 Access Points for Amazon FSx are now available in additional AWS Regions
17. [17] Amazon EKS Auto Mode and Karpenter now support Dynamic Resource Allocation