Home› Blog› AWS Weekly Intelligence #9 - 28 September-2 Octobe…
AWS Weekly Intelligence AWS

AWS Weekly Intelligence #9 - 28 September-2 October 2026

Oct 3, 2026 22 min read
Verified against current vendor documentation on 3 October 2026. Pricing, limits and API behaviour were checked against the official docs on that date. Cloud services change fast — if you are reading this much later, treat the specifics as a starting point and re-check the linked sources.
On this page

The week in one paragraph

79 announcements across five working days, weighted heavily towards security and governance, with Wednesday 30 September carrying 25 of them on its own. Five deep-dives went out, one per news day, and three of them landed on the same failure shape without being chosen for it: a configuration that is absent rather than wrong. S3 Vectors was returning fewer results than asked for because a filter arrived too late. A GuardDuty declarative policy silently drops every feature a Region override forgets to re-list. An ECS linear deployment accepts two values that are individually valid and together exceed a third documented timeout. None of the three produces an error.

The other through-line is that AWS spent the week wiring governance into places it had not reached. GuardDuty got declarative policies, Security Hub got remediation plans and absorbed GuardDuty Runtime Monitoring into its Threat Analytics plan, Secrets Manager got posture recommendations, IAM Identity Center widened its multi-Region footprint, and AWS Health started tracking software version lifecycles. Five separate services, one direction of travel.

Covered in depth

One post per news day. The column that matters is the last one — what the announcement did not say.

News date Post What the announcement left out
28 Sep #42 — The air gap is that they are not in your account The backups sit in an AWS Backup service-owned account, which CloudTrail reports as shared outside your organization. Always compliance-mode locked, 7-day minimum retention, KMS key immutable after creation, and RAM sharing to individual account IDs only.
29 Sep #43 — The third agent runtime Bedrock Agents is now Bedrock Agents Classic and closed to new customers, while AgentCore Harness already offers a managed agent loop that already reaches OpenAI. Three overlapping runtimes, one being retired quietly.
30 Sep #44 — The filter ran after the search The cutover is a date, not a flag: buckets created before 30 September 2026 are CLASSIC, and the docs now admit that on a CLASSIC index "queries with filters may return fewer than top K results". The 5x figure is a confession about the old behaviour.
1 Oct #45 — A Regional override replaces the default A Region block "fully replaces the default for that Region" rather than extending it, so an override written to disable one feature drops everything it did not re-list. Omitting the default block forfeits automatic coverage of future Regions, and detaching departs from the general declarative-policy rollback rule.
2 Oct #46 — The most cautious deployment the API accepts cannot finish stepPercent goes to 3.0 and stepBakeTimeInMinutes to 1440; together they are 33 days against a 30-day overall deployment timeout. CloudFormation caps the whole deployment at 36 hours. And the circuit breaker is rolling-update only, so these strategies roll back on alarms or not at all.

Two Architecture Series posts this week sit directly against the week's news rather than beside it. #35's seven-gate evaluation order is the permissions counterpart to GuardDuty's declarative policies — one denies API calls, the other enforces configuration — and #71, published this morning on AWS Private CA, is the service behind Friday's AgentCore Gateway private-TLS announcement.

What else shipped, by domain

Security and governance

The densest cluster of the week, and all of it unwritten. Security Hub remediation plans arrived to "prioritize and fix security exposures", and separately GuardDuty Runtime Monitoring is now included in the Security Hub Threat Analytics plan — a packaging change worth reading carefully alongside #45, because runtime_monitoring is one of the eight keys a GuardDuty declarative policy governs, so enablement and billing now arrive from two different directions. Secrets Manager gained actionable posture recommendations in the console. IAM Identity Center extended multi-Region support to more Regions. AgentCore Gateway began accepting private TLS certificates for VPC endpoints. S3 Object Lock variable retention with event holds reached GovCloud (US).

Data and analytics

DynamoDB filtered export to S3 is the notable one — exporting a subset rather than the whole table changes the economics of feeding a lake from an operational store. Redshift gained cross-Region queries for data lakes, Glue Data Catalog added table optimization, statistics and crawlers for Apache Iceberg V3, EMR Serverless got terabyte-scale shuffle on Serverless Storage, and S3 Tables raised the limit to 100 table buckets per Region per account.

Databases

Aurora DSQL partial indexes — an index over a subset of rows, which "improves query performance and lowers index storage cost" — is the most interesting, because DSQL's pricing model makes storage decisions visible in a way provisioned engines do not. DocumentDB had a strong week with retryable writes plus five MongoDB aggregation stages and change stream capabilities in 8.0.2. Aurora and RDS for PostgreSQL shipped the usual minor-version sweep, and Aurora Serverless gained instant 16-ACU scaling.

Compute and containers

EKS and EKS Distro support Kubernetes 1.37, which starts the usual upgrade clock. EC2 future-dated Capacity Reservations can now have their start dates postponed — small, and genuinely useful if you have ever had a reservation activate against a project that slipped.

AI and agents

Besides Bedrock Managed Agents (covered in #43): GPT-6 Astra gained UltraFast mode on Bedrock, GPT-6.1 Sol reached GA, Grok 4.7 and Claude Sonnet 5.5 became available, Claude model availability expanded to India, South Korea and Singapore, and the Well-Architected Agent entered preview. The AWS MCP Server reached six additional Regions.

Operations and cost

AWS Health's version catalog is the sleeper item of the week: "a centralized source of lifecycle information for software versions across AWS services". Anyone who has tried to build an end-of-life inventory across RDS engine versions, EKS versions, Lambda runtimes and managed Kafka by scraping documentation should look at this before building it again. Systems Manager documents can now be shared through AWS RAM, ElastiCache for Valkey gained OpenTelemetry metrics and detailed monitoring, Transfer Family supports custom CloudWatch log groups for managed workflows, and Budgets added email verification for notification subscribers.

Security bulletins

Three on 1 October, of which one is the kind this series keeps finding: CVE-2026-104002, fail-open error handling in the data masking utility in Powertools for AWS Lambda (Python) — a masking utility that fails open does not mask. Also CVE-2026-97662, argument injection in the security-agent-mcp-server diff scan, and CVE-2026-104020, uncontrolled recursion in the Ion reader in Amazon Ion Python.

What I would act on

1. Audit existing GuardDuty Region overrides today. This is the only item of the week that can already be configured wrong rather than becoming wrong later. If you piloted declarative policies and wrote a Region block to adjust one feature, that block replaced the default for that Region and silently dropped every feature it did not re-list. Check the account status report, not the policy document — the policy cannot show you a Region it never asserted anything about.

2. Multiply your ECS linear deployment settings. If any service uses the LINEAR strategy with a small step percent and a long step bake time, compute steps × bake time and compare it against 36 hours if CloudFormation deploys that service. Nothing warns you — both values are inside their documented ranges, and the failure is a timeout that rolls back a revision already serving most of your traffic.

3. Check whether your S3 Vectors buckets are CLASSIC. Every vector bucket created before 30 September 2026 is, and on a CLASSIC index a selective filter can return fewer results than you asked for with no error. If you run retrieval over a filtered corpus, this has been quietly costing you recall. The fix is a new bucket, not a setting.

4. Look at the AWS Health version catalog before your next end-of-life exercise. Low urgency, high leverage. Software lifecycle tracking across AWS services has been a bespoke scraping job in most organisations; a first-party catalogue changes whether that job needs to exist.

One thing I would specifically not rush: GuardDuty Runtime Monitoring moving into the Security Hub Threat Analytics plan. It is a packaging change, the cost question depends on which plan you are already on, and it interacts with declarative-policy enablement in a way worth working out deliberately rather than in the week it shipped.

Complete inventory

All 79 announcements from the AWS What's New feed for 28 September to 2 October 2026, newest first, with AWS's own one-line summary for each. Built from the raw feed rather than summarised, and every link verified to return HTTP 200 at publication.

DayAnnouncements
Monday 28 September10
Tuesday 29 September19
Wednesday 30 September25
Thursday 1 October16
Friday 2 October9
Total79

Friday 02 October — 9 announcements

Thursday 01 October — 16 announcements

Wednesday 30 September — 25 announcements

Tuesday 29 September — 19 announcements

Monday 28 September — 10 announcements

Looking for one service rather than a whole week?

Every AWS, Azure and Google Cloud announcement is browsable by service and date, each linked to the vendor’s own page — and the other weekly roundups are collected in one place.

Browse announcements →
Topics: AWS Weekly Intelligence AWS

Comments

How was your experience?
Your feedback helps improve this site.
PoorExcellent