Home› Blog› AWS Weekly Intelligence #8 - 21-25 September 2026
AWS Weekly Intelligence AWS

21-25 September 2026

Forty announcements across five working days. Three of them moved a decision across a boundary — from the bus owner to the consumer, from the producer to the service, from the developer to the telemetry pipeline — and a run of CVEs landed in code that runs inside your own process.

Sep 26, 2026 16 min read
AWS Weekly Intelligence #8
Jayanth Katta 26 Sep 2026 Every announcement in the window, ranked and linked
On this page

The week in one paragraph

Forty announcements, distributed unevenly: Monday produced two, Tuesday eleven. The volume was in AI availability — GPT-6 Sol and Luna reached general availability on Bedrock, Claude Opus 5.5 arrived on AWS and on GovCloud, and Bedrock Managed Knowledge Base gained Salesforce and Zendesk connectors. The substance was elsewhere. Three announcements each moved a decision from one party to another, which is a more consequential kind of change than a feature: EventBridge moved event routing from the bus owner to the consumer, Kinesis moved partition-key selection from the producer to the service, and CloudWatch Omni moved every prompt and tool call into the observability plane. Running underneath all of it was a poor week for drivers and hypervisors — a remote code execution in the Redshift JDBC driver, an out-of-bounds write in Firecracker, an RCE in pgcollection, and issues in FreeRTOS-Plus-TCP, the Athena ODBC driver and the AWS C event stream decoder. Every one of those runs inside a process you own rather than behind an API you call.

Covered in depth

Five daily posts published against this window, and all five cover announcements made inside it — so unusually, there is nothing to reconcile across the boundary this week.

Post Announcement What the announcement did not say
#37 — You can now watch it fail in real time ECS real-time deployment observability in the console (Mon 21) Observability of a rollout is not control of one; the console view and the circuit-breaker are different features.
#38 — Every prompt is now telemetry CloudWatch Omni GA (Wed 23) It captures every prompt, and CloudWatch Logs masking applies only at ingestion — so the data protection policy has to exist before the first trace.
#39 — The partition key was doing two jobs Kinesis Service-Managed Partition Keys (Wed 23) The partition key was also the ordering guarantee, which is why the feature is scoped to workloads "where record ordering is not required".
#40 — A hundred buses became five EventBridge relaunches event buses (Thu 24) The default bus allowance fell from 100 to 5, and subscriber creates are serialised to one at a time across every account sharing the bus.
#41 — The connection string is executable CVE-2026-8178, Redshift JDBC driver (Fri 25) Connection-URL options documented to take class names and file paths, so the CVE is a documented surface failing rather than a new one.

Three announcements moved a decision across a boundary

This is the week's real pattern, and it is worth naming because a feature you can ignore and a reassignment of ownership you cannot look identical in a release note.

EventBridge moved routing from the producer to the consumer. Under the classic bus, a consumer wanting events filed a request against somebody else's rules. The relaunched bus gives consumers their own subscriber resources — and the quota page is where the consequence shows: five buses instead of a hundred, ten thousand subscribers each. The bus stopped being a fence between teams.

Kinesis moved partition placement from the producer to the service. Service-Managed Partition Keys distributes by available warm capacity instead of by MD5 hash of a key you chose. It also takes the per-key ordering guarantee with it, because those were always the same mechanism.

CloudWatch Omni moved prompts into telemetry. Capturing every prompt, model call and tool invocation is exactly what agent debugging needs. It also relocates user-typed text into a plane designed to be widely readable.

In all three, the useful question is the same: who owns the decision now, and does anything downstream still assume the old owner did?

A bad week for code that runs inside your process

Six security bulletins landed on components that are not services. They are libraries and drivers linked into your own runtime, which changes both the blast radius and how you find them.

  • CVE-2026-8178 — remote code execution via unsafe class loading in the Amazon Redshift JDBC driver, below 2.2.2, fixed in 2.2.2. The broadest reach of the six, because JDBC drivers arrive bundled inside BI tools, Glue connections and fat JARs.
  • CVE-2026-96883 — type confusion allowing remote code execution in pgcollection.
  • CVE-2026-5747 — out-of-bounds write in the Firecracker virtio-pci transport. A hypervisor boundary.
  • CVE-2026-5190 — stack buffer overflow in the AWS C event stream streaming decoder.
  • FreeRTOS-Plus-TCP — MAC address validation bypass and an ICMP echo reply integer underflow.
  • Athena ODBC driver and AWS Ops Wheel (CVE-2026-6911, CVE-2026-6912) — issues published the same week.

The common instruction across these bulletins is the one people skip: patch forked and derivative copies too. A vendored, shaded or relocated library does not upgrade when the dependency does, and it does not appear in a dependency listing under the name you would search for.

The model shelf refilled

Tuesday carried four model announcements. OpenAI GPT-6 Sol and GPT-6 Luna reached general availability on Bedrock. Claude Opus 5.5 became available on AWS and, separately, on AWS GovCloud (US) — the GovCloud entry is the more interesting of the two, because frontier models reaching a government partition is usually a slower path than this.

Bedrock Managed Knowledge Base added Salesforce and Zendesk as native data source connectors. Worth pairing with the observation from the architecture series this week: those connectors pull text written by customers and support agents into a retrieval corpus, and content outside input tags is not processed by Guardrails. A new connector is a new authorship boundary.

SageMaker HyperPod Inference Gateway arrived as a Kubernetes-native, GPU-aware routing layer deployed as a single EKS add-on with, per the announcement, zero application changes. That is the self-hosted counterpart to the week's managed-model news.

Sovereignty, and where things are allowed to run

A consistent thread, and mostly regional expansion rather than new capability. EC2 M8i, R8i and C8i families (and their flex variants) all reached the AWS European Sovereign Cloud (Germany), as did Lambda durable functions. Amazon EVS came into scope for FedRAMP Class C across all US Regions.

Read alongside CloudWatch Omni's three-Region availability from the same week, the shape is familiar: new capability arrives in a handful of Regions and the compliance envelope follows later. Anyone with residency obligations is choosing between the newest service and the Region it is required to run in, and this week did not change that.

Post-quantum showed up twice, quietly

RDS for PostgreSQL added post-quantum TLS key exchange, which was the announcement. The second appearance was not announced at all: the Redshift JDBC driver's option list documents preferPQ, controlling whether the driver advertises post-quantum hybrid key-exchange groups in the TLS 1.3 ClientHello, available in driver versions 2.2.8 and later.

That option also documents the failure mode worth knowing before you enable any of this: it exists so you can turn it off when “a network middlebox blocks TLS handshakes because it does not recognize the post-quantum key-exchange groups the driver offers.” Post-quantum migration is going to surface as inspection appliances rejecting handshakes, not as cryptographic failure.

Quieter, and worth knowing

DynamoDB multi-Region strong consistency reached five more Regions and, more significantly, cross-continent configurations — strong consistency across continental distance is a latency conversation, not just an availability one.

IAM outbound identity federation added interface VPC endpoints for OIDC discovery, which removes an internet egress requirement from a federation path. Network Security Manager reached GA, in us-east-1 only. Route 53 Resolver became generally available on second-generation Outposts racks.

Amazon EMR had a busy week: 7.14 released, Spark Connect for interactive workloads on EMR on EKS, IPv6 EKS cluster support, and the introduction of Long Term Support releases starting with Apache Spark 4.1 — that last one is the most useful for anyone who has been pinned to an EMR version by an upgrade they could not schedule.

Billing and Cost Management shipped a ListBillingViewSegments API returning account billing context over a period, and Billing Transfer gained automatic billing group creation for two-level transfers. Glue Data Quality now generates rule recommendations in seconds. Elastic Disaster Recovery added Graviton-based source servers. Transcribe added customer-managed KMS keys for custom vocabularies and language models. DataSync gained an account-wide monitoring dashboard. Security Hub AI Inventory extended discovery to self-hosted instances in Azure.

What I would act on

1. Inventory every Redshift JDBC driver and get to 2.2.2. This is the week's only item with a published remote code execution against something most data platforms link into a JVM holding a real IAM role. The upgrade is trivial; the inventory is not, and it has to include vendored and shaded copies. Do this one first.

2. Write the CloudWatch Logs data protection policy before you create an Omni space. Masking applies at ingestion and is not retroactive, so the ordering is the only part of this that cannot be fixed afterwards. If Omni is on anyone's evaluation list for next quarter, the policy is this quarter's work.

3. Check whether anything downstream of a Kinesis stream assumes per-key ordering before a producer stops sending partition keys. The feature is adopted by a library upgrade, which means it can arrive as a dependency bump, and the failure is silent: out-of-order processing surfaces as a wrong total, not an error.

4. Make subscriber creation on the new EventBridge bus retry ConcurrentModificationException. One create or delete in progress per bus, across every account it is shared with, is expected behaviour on a shared bus and will look like a broken pipeline the first time two teams deploy together.

5. Set SSLMode=verify-full and SSL_Insecure=false on Redshift JDBC connections while you are already touching them. Both are weaker by default than AWS's own recommendation, and you will not have a better excuse to change them than a driver upgrade you are doing anyway.

Complete inventory: all 40 announcements

Every AWS What's New announcement in the window, by day, with AWS's own summary. Built from the raw feeds and link-validated rather than summarised.

Friday 25 September — 10 announcements

Thursday 24 September — 11 announcements

Wednesday 23 September — 6 announcements

Tuesday 22 September — 11 announcements

Monday 21 September — 2 announcements

40 announcements, 40 unique links, all validated. Security bulletins are listed in the driver section above rather than here, because they are published on a separate feed from What's New.

Official AWS references

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