Home› Blog› AWS Weekly Intelligence #6 - 7-11 September 2026
AWS Weekly Intelligence AWS

AWS Weekly Intelligence #6 - 7-11 September 2026

Sep 12, 2026 17 min read
Verified against current vendor documentation on 12 September 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.

The week in one paragraph

Four working days, because Monday was Labor Day and the What’s New feed returned nothing at all. They still produced 53 announcements, and the most consequential item was not an announcement — it was a security bulletin published on Thursday.

Read together, the week has a shape: one urgent CVE in the service teams adopted specifically to be safer, one compliance primitive that closes a gap records policies have had since Object Lock shipped, Lambda acquiring a second compute model with genuinely different economics, and two attribution tools — both of which only work if they were installed before you needed them.

Covered in depth

PostSubjectThe finding
Daily #28 S3 Object Lock variable retention Releasing a legal hold ends protection; releasing an event hold starts it. Same verb, opposite outcome.
Daily #29 Systems Manager unmanaged-node diagnosis OS issues are read from console output and cannot be auto-remediated, because fixing them needs the access you lost.
Daily #30 CloudWatch NHI for TGW peering 0 means no degradation observed, not healthy — and it is delayed by hours after any monitor change.
Daily #31 CVE-2026-89049, SSM Agent SSRF IMDSv2 binds the session token to the instance. It does not bind the credentials, which is why they travel.
Architecture #46–#50 DNS egress, flow logs, logging cost, cost allocation, commitment sharing Network Firewall has no visibility into Resolver queries; a flow log format cannot be changed retroactively; split charge rules never reach the Cost and Usage Report.

The bulletin that should reorder your Monday

CVE-2026-89049 is a server-side request forgery in Session Manager port forwarding in SSM Agent, affecting every version below 3.3.4851.0. The mechanism, in AWS’s words, is improper validation of equivalent address representations — a denylist of forbidden remote destinations compared the spelling of an address rather than the address, so an equivalent representation of a link-local address passed.

What sits at that link-local address is the instance metadata service. So an authenticated user with port-forwarding permission could obtain the managed instance’s temporary IAM role credentials and act with that role’s permissions from outside the instance.

The reason this deserves the top of the page rather than a line in the inventory: Session Manager port forwarding is the recommended way to retire bastion hosts. The permission is therefore granted broadly and deliberately, to exactly the people you would not want holding an arbitrary instance role.

Two speeds of response. AWS’s interim measure — scoping ssm:StartSession so untrusted principals cannot invoke AWS-StartPortForwardingSessionToRemoteHost — applies centrally today with no instance access. The upgrade past 3.3.4851.0 has to reach every managed node and will leave a tail.

Friday added four more, all in developer tooling rather than infrastructure: CVE-2026-89332, sensitive workspace data exfiltration in Kiro IDE via agent-written workspace configuration; CVE-2026-89090, a denial of service in the event stream header decoder in the AWS SDK for Go v2; CVE-2026-18061, an XXE in the AWS Advanced JDBC Wrapper; and path traversal plus OS command injection in projen.

A retention clock that starts when something happens

Object Lock gained variable retention. Until now it offered a retention period whose clock starts at write time, and a legal hold with no clock at all. Neither expresses the sentence real records policies are written in: retain for seven years after the account closes.

Now you place an event hold with a duration, and when you release it S3 fixes the retain-until-date at release time plus that duration. Free, every Region including China, and applicable to individual objects, bucket defaults, or an existing estate through S3 Batch Operations.

The trap is that it superficially resembles a legal hold and behaves in the opposite direction. A legal hold has no expiration and stays until explicitly removed, so removing it ends protection immediately. Releasing an event hold begins it. Any runbook that treats “release the hold” as a single concept will get one of the two backwards.

Lambda now has two compute models with different physics

The headline was a 90-minute function timeout, a 6x increase on the 15-minute limit. The qualifications matter more than the number: it applies to asynchronous and event source mapping invocations, on Lambda Managed Instances only. Synchronous invocations keep 15 minutes everywhere.

The limit could move because the meter moved. On Managed Instances you pay EC2 instance pricing plus a 15% premium, and do not pay separately for the execution duration of each request. Duration stops being the billing unit, so a long invocation stops being priced like one. Graviton5 support and a durable-functions integration landed in the same week, which reads as a compute model being filled out rather than a feature being added.

What you give up is scale-to-zero. The default type charges nothing when idle; a capacity provider bills for instances whether or not they are working.

Two tools for “is it us or is it AWS”, both needing a head start

CloudWatch extended its network health indicator to paths crossing a Transit Gateway inter-Region peering connection. It is a binary metric: 100 means degradation was observed in the AWS-controlled path, 0 means none was observed. Read that second value carefully — it is not “healthy”, and for a peered path the indicator only covers the AWS path up to the peering connection, so a fault beyond it also reads 0.

Systems Manager expanded its unmanaged-node diagnosis by six categories: IAM permissions, agent version, instance status checks, operating system configuration, Default Host Management Configuration, and hybrid activation. The best detail is an honest one — OS issues are diagnosed by reading the instance console output, and cannot be auto-remediated, because remediation requires direct access to the instance operating system. The agent is down because the disk is full; the full disk is why the agent cannot clear it.

Both share a property worth naming. The NHI is delayed by a few hours after any monitor or probe change while anomaly detection builds a baseline. Neither is something you stand up during the incident it would have explained.

Quieter, and worth knowing

  • API Gateway now supports mutual TLS for backend integrations — the other half of a story where only the client leg had mTLS.
  • ECS added ecs:task-cpu and ecs:task-memory condition keys on RunTask and StartTask, so CPU and memory ceilings can be enforced through IAM rather than convention. All Regions, no cost.
  • Second-generation single-rack Outposts, for anyone whose edge footprint is a capacity-planning exercise.
  • Storage Gateway gained FIPS-compliant private connectivity for S3 File Gateway.
  • EBS Volume Clones can now copy across accounts.

What I would act on

PriorityActionWhy now
Before Monday Scope ssm:StartSession by document; inventory SSM Agent versions against 3.3.4851.0 The IAM scope is central and immediate. The upgrade is the real fix and will take longer.
This sprint Patch the Go SDK v2, JDBC Wrapper and projen findings in your build pipelines Developer-tooling CVEs reach production through the supply chain, not the fleet.
This quarter Identify retention policies written as “N years after X” and model them with event holds Anything currently counting from write time is over-retaining by the whole life of the relationship.
Before you need it Deploy NHI monitors on production cross-Region paths The indicator is unavailable for hours after creation. Standing it up mid-incident answers nothing.

Complete inventory: all 53 announcements

Every AWS What’s New announcement published between 7 and 11 September 2026, grouped by day. Monday produced none — it was a US public holiday — and that absence is real rather than a gap in collection.

Friday 11 September — 7 announcements

Thursday 10 September — 17 announcements

Wednesday 09 September — 15 announcements

Tuesday 08 September — 14 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