August produced 244 announcements across 22 working days, and they were not evenly distributed in subject. Two thirds of the month was routine — Region expansions, instance families, database minor versions — and the remaining third clustered tightly around three things: identity and access, operational recovery moving into the platform, and agent infrastructure acquiring the governance controls that ordinary infrastructure already has.

The month ended with a security bulletin that needs action rather than reading. CVE-2026-81849 is a path traversal in the Amazon SSM Agent, reachable by a caller whose permissions were deliberately narrowed, and the affected versions span most of the agent's release history. A second bulletin on 31 August covers remote code execution in the OpenSearch SQL plugin. Both are in the action list below.

Nothing in the month changes our architecture direction. Several things make work we already planned cheaper: AWS Glue 6.0 arrived with a documented 30% price reduction, Elastic Disaster Recovery gained recovery ordering that replaces a runbook, and AWS Certificate Manager added a way to switch existing certificates from email to DNS validation without reissuing them — which materially reduces the cost of a migration we have to complete by September 2027 regardless.

Five items with a clear owner and a date. The first two are this week.

Operational recovery is moving into the services

The clearest pattern of the month, and it appeared five separate times. EC2 gained application-level status checks so Auto Scaling can replace an instance whose OS is healthy and whose application is not. ECS began emitting agent connectivity health events and can drain and replace impaired container instances on its own. EKS gained a managed certificate authority rotation process. Elastic Disaster Recovery gained Recovery Plans, which sequence a multi-server recovery with waits and approvals. AWS Batch can now run on ECS Managed Instances, with AWS owning AMI updates and host patching.

Each of those replaces something a platform team currently writes and maintains. The consequence for us is a review rather than an adoption programme: for each piece of custom recovery automation we run, the question is now whether the service does it, and whether ours is better enough to keep.

Identity was the month's dominant theme

Lambda functions gained full IAM resource-based policies, with multiple principals and the complete condition-key surface. EKS clusters can now trust several external OIDC providers at once. Cognito added an administrative operation to reset a user's TOTP configuration, and later machine-to-machine authorization without a user pool domain. IAM Roles Anywhere shipped a Java credential provider, removing the last practical reason to leave long-lived access keys on hybrid Java hosts.

The Cognito TOTP reset is the one worth a second look internally. Resetting a second factor is operationally indistinguishable from bypassing it, so that permission belongs in the same review as password reset — and in most estates it is not there yet.

Agent platforms started acquiring the controls ordinary infrastructure has

AWS Agent Registry reached general availability on 31 August, giving agents and MCP servers centralised discovery and governance. Bedrock AgentCore added temporal authorization policies and gateway rate limiting, and its Memory service gained flexible namespaces and fine-grained access control through Cedar policies — per-tenant isolation that teams have been writing by hand.

Read alongside the month's Strands Agents Tools advisory, in which a consent prompt could be skipped by passing a parameter through another tool, the direction is unambiguous: the agent's own reasoning is not the security boundary. Identity, gateway policy and sandboxing are. Any internal agent work should assume that from the start rather than retrofit it.

Operational and analytical data keep converging

DynamoDB gained native vector search, so embeddings can live beside the operational records they describe. Redshift added Apache Iceberg v3 tables and long-term system table retention through S3 Tables. Glue 6.0 brought Iceberg v3 with the VARIANT type and deletion vectors. The practical effect is that the default answer to “we need a second system for this” is worth re-testing; several of the reasons for one were removed this month.

Dated items only. Everything here was read from AWS's own documentation.