AWS News - 2026-10-21

Today, AWS launched the public beta of TorchNeuron, an open source native PyTorch backend for AWS Trainium. Trainium is a portfolio of AI chips designed to deliver high performance and cost efficiency for deep learning AI training and inference workloads. TorchNeuron enables AI scientists and researchers to experiment and build models with PyTorch natively on Trainium.

With TorchNeuron, developers can enable their PyTorch code on Trainium simply by switching to .to('neuron') while maintaining full PyTorch compatibility with eager execution, torch.compile, and standard PyTorch APIs. Additionally, TorchNeuron integrates with Neuron Kernel Interface (NKI) to enable custom operators and performance optimization for Trainium hardware.

TorchNeuron is available as open source through the TorchNeuron github repo for Amazon EC2 Trainium2 instances and Amazon EC2 Trainium3 instances.

To learn more about TorchNeuron and AWS Trainium accelerators, visit the AWS Neuron page.

Amazon DynamoDB global tables now support disassociating a replica without deleting its table or data. Global tables provide fully managed, multi-Region, multi-active replication. You can now stop replication to and from a replica and continue using the same table as a single-Region table.

An account can leave a shared multi-account global table and continue using its table independently. You can also stop replication to a Region while keeping its table for local workloads, validation, or audit, without creating a replacement table or copying data.

DynamoDB supports disassociating replicas in all AWS Regions for same-account and multi-account global tables that use multi-Region eventual consistency (MREC).

To get started, use the AWS Management Console, AWS CLI, DynamoDB API, or AWS CloudFormation. For more information, see Disassociating replicas in the Amazon DynamoDB Developer Guide.

Today, Amazon SageMaker AI announces the ability to configure the sequence length used during your serverless model customization job. Amazon SageMaker AI serverless model customization supports fine-tuning for many open-weight models, including models from popular families including gpt-oss, Gemma, Llama, Nemotron, and Qwen. Previously, each fine-tuning recipe ran at a fixed sequence length that customers could not modify. Customers sometimes need to modify these parameters to achieve better results for their use case. Customers also asked for longer sequence lengths than the static defaults we supported on some models.

Now, you can set sequence length parameters as part of the customization job's hyperparameters, anywhere up to the published maximum for your chosen model and technique. You can set the dataset_max_len hyperparameter for SFT and DPO jobs or set max_prompt_length and max_response_length hyperparameters for RFT jobs. This gives you more control to match your training data characteristics, whether you are working with short question-answer pairs or multi-page documents. Published maximums are up to 16K tokens for many models depending on model and technique.

Serverless model customization on SageMaker AI is available in US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland). To get started, navigate to the Models page in Amazon SageMaker Studio to launch a customization job, use the SageMaker Python SDK or use our Agent skills with your choice of coding assistant. To learn more, and see the list of models and supported sequence lengths per technique, see the SageMaker AI documentation.

AWS Network Firewall now supports IP validation, extending domain-based filtering beyond HTTP/HTTPS to protocols such as SFTP and SSH. This new capability validates that a connection's destination IP address was resolved through your VPC's Amazon Route 53 VPC Resolver, closing security gaps where clients could bypass domain rules.

This feature enables Network Firewall customers to implement domain filtering rules for any protocol, not just HTTP/HTTPS, providing more flexible security policies than IP-based rules alone. It simplifies protection against Server Name Identifier (SNI) spoofing by validating the IPs resolved by the Route 53 VPC Resolver against the SNI, ensuring traffic only reaches legitimately resolved destinations. Organizations can now enforce consistent domain-based security policies across all outbound traffic types, from web applications to file transfer protocols, without complex workarounds. Additionally, Route 53 Resolver DNS Firewall can be used with your VPC Resolver for extended DNS security.

Resolved IP validation is available in all AWS commercial regions where AWS Network Firewall and Route 53 Resolver DNS Firewall are available, at no additional charge. For a full list of supported regions, visit the AWS Capabilities by Region page.

To learn more, visit AWS Network Firewall product page and service documentation.

AWS Transform for migrations now gives you three ways to bring your network to AWS. In addition to migrating and modernizing your existing network, you can build a new network from scratch or convert your on-premises access rules into optimized security groups and deploy them into existing VPCs. Whichever path you choose, you arrive on AWS with a secure and scalable network foundation.

On-premises networks accumulate years of definitions and access rules that you may not want to carry to AWS as-is. For a clean start, AWS Transform builds a new network aligned to AWS best practices. To bring your access rules with you, AWS Transform converts them into security groups and modernizes them as it goes, detecting stale rules and recommending optimizations, then deploys the groups into your existing VPCs without redeploying your network stack.

These capabilities allow you migrate to any AWS commercial or GovCloud Region.

To get started, see Migrate networks. To learn more, visit AWS Transform for migrations.

Amazon S3 server access log delivery to Amazon CloudWatch Logs is now available in the AWS GovCloud (US) Regions, AWS European Sovereign Cloud, and AWS China Regions. With your access logs in CloudWatch Logs, you can query them as they arrive, set alarms, and encrypt them with AWS Key Management Service (AWS KMS) keys. This new delivery path complements the existing delivery of server access logs to S3 general purpose buckets, giving you more flexibility in how you monitor and analyze access to your data.

With delivery to CloudWatch Logs, you can set alarms on error rates, monitor traffic patterns, investigate access incidents, and correlate S3 access activity with the rest of your operational data.

S3 server access log delivery to CloudWatch Logs is now available in all AWS Regions. Amazon CloudWatch Logs charges for vended logs apply. For pricing details, visit Amazon CloudWatch pricing. To learn more, visit the Amazon S3 User Guide.

Amazon Connect Customer now offers two different layouts that optimize screen space on the agent workspace. Contact center agents spend hours each day in the agent workspace, which includes all of the contact controls and applications they need to resolve customer issues. Agents who handle different channels and workflows want the workspace arranged differently, so that they can quickly read, respond, and access their tools. Now, administrators can choose the best layout for each team, maximizing agent efficiency and satisfaction.

On the Modern layout, contact controls are at the top of the workspace, maximizing screen space for the applications beneath during voice calls. When agents handle chat, email, and task contacts on the Modern layout, they now have a taller, wider area to read and respond. The Classic layout offers a familiar experience for agents already using the agent workspace, now with contact controls that expand and contract by channel and a resizable agent assist application. Administrators can assign the Modern layout to a team migrating from another agent desktop with controls at the top, while teams already familiar with the agent workspace can use the Classic layout.

You can use the new layouts for the Amazon Connect Customer agent workspace in the following AWS regions: US East (N. Virginia), US West (Oregon), Africa (Cape Town), Asia Pacific (Seoul), Asia Pacific (Singapore), Asia Pacific (Sydney), Canada (Central), Europe (Frankfurt), and Europe (London).

To learn more and get started, see the agent workspace in the Amazon Connect Customer administrator guide.

Try Amazon ECS Gateway, a new account-level resource that simplifies ingress networking for containerized workloads by dynamically provisioning and managing the necessary infrastructure on your behalf, with all resources created directly in your account.

2025 年 9 月以降に相次ぐ、npm Registry を悪用した Shai-Hulud などのサプライチェーン攻撃を踏まえ、AWS Well-Architected フレームワークのセキュリティの柱に沿ったパッケージ利用者向けのベストプラクティスを紹介します。一時的な認証情報と最小権限による漏えいリスクの低減、AWS Signer によるアーティファクト署名などの多層防御を取り上げます。さらに、AWS CodeArtifact による依存関係の一元管理、Amazon Inspector による依存関係スキャン、AWS CloudTrail と Amazon GuardDuty によるモニタリングも解説します。

npm や PyPI のパッケージは、公開直後の数時間に最もサプライチェーンのリスクが高くなります。この記事では、Amazon Linux 2023 で npm の min-release-age と pip の uploaded-prior-to を使用して 1 日の依存関係のクールダウンを設定し、予期しないパッケージのインストールを防ぐ方法を紹介します。また、npm audit と pip-audit で特定したパッケージに限ってクールダウンをオーバーライドし、セキュリティ修正を適用する方法も説明します。

Kiro などの AI コーディングエージェントは開発を加速させる一方で、プロンプトインジェクションや過度に緩い設定、サプライチェーンリスク、スコープクリープなどのリスクをもたらします。本記事では、IDE でエージェントの出力を方向付ける作成時のコントロールと、パイプラインで検証するビルド時のコントロールの 2 つの柱からなるアプリケーションセキュリティのフレームワークを紹介します。決定論的な緩和策、非決定論的な緩和策、人間によるレビューの組み合わせ方に加えて、Kiro、AWS CodeBuild、AWS CodePipeline を使用した実装例も解説します。

OpenAI ChatGPT Codex のモデル推論を Amazon Elastic Container Service (Amazon ECS) 上の LiteLLM ゲートウェイ経由で Amazon Bedrock にルーティングし、生成 AI コーディングエージェントをエンタープライズ向けに一元管理する方法を紹介します。AWS Fargate へのデプロイ手順、スコープ付きキーによる予算とレート制限の適用、Responses API の互換性検証を解説します。AWS IAM Identity Center による直接アクセスや Portkey を使用する代替アクセス経路も取り上げます。

AWS DevOps Agent のような自律型 AI エージェントは、調査中に呼び出す API、回数、対象スコープを動的に判断します。そのため、調査能力だけでなく、実行前にコストを把握して制御する仕組みも重要です。Investigation Cost Guardrail スキルを利用すると、最初の課金対象クエリを実行する前に見積もりを確認し、続行するかどうかを判断できます。本記事では、このオープンソーススキルが AWS サービスやサードパーティーツールの操作を分類し、リージョンごとのライブ料金とデータ転送コストを取得する仕組みを解説します。また、Proceed（続行）、Warn（警告）、Halt（停止）の判定ゲートを、3 つの具体的な調査シナリオとともに紹介します。

本日、Kiro IDE と CLI にメモリ機能を導入します。Kiro は、あなたが好む作業の進め方、リポジトリ固有のコマンドや規約、受けた訂正、デバッグで得た知見、すでに除外したアプローチを覚えておけるようになります。それらが再び関係する場面になると、Kiro はあなたのコンテキストを引き継ぎます。コンテキストを説明し直したり同じ訂正を繰り返したりする時間が減り、Kiro は言われなくてもあなたの好みに沿って動けるようになり、使い込むほど役に立つ存在になります。

Kiro のファイル読み取りツールと検索ツールの説明に、ハーネスの動作について事実と異なる主張をする一文がありました。この一文は数百のセッションで 61,742 回モデルに届いていました。何の役にも立っておらず、誰かがチケットを起票することもまずなかったでしょう。こうした問題は見つけるのが難しく、修正の良し悪しを判断するのも同じくらい難しいものです。たとえば私たちは、より多くのタスクに合格し、使用トークンも少ない修正をあやうくリリースするところでした。結果をホストごとに分けてみると、改善のほぼすべてが、エージェントがいつもはまり込むループからたまたま抜け出した 1 回の実行によるものだとわかったのです。そのホストを除くと、その修正は遅く、コストが高く、精度も下がっていました。自分の修正を自分で採点するループなら、これをリリースしていたでしょう。

Bulletin ID: 2026-140-AWS
Scope: AWS
Content Type: Important (requires attention)
Publication Date: 10/20/2026 13:00 PM PDT

Description:

Kiro Crew is an open-source personal AI agent that can be reached from a web dashboard, a CLI, or messaging channels such as WhatsApp, and that runs tools on the operator's machine.

We identified CVE-2026-112380, an authorization issue in the WhatsApp channel and the internal subagent run-control endpoints. In certain configurations where a WhatsApp group was added to an operator's allowed groups, or where the direct-message policy was set to open, a member of that WhatsApp group could send instructions to another session's subagent. The subagent then executed those instructions on the operator's machine under that session's approval policy, and the resulting audit trail was attributed to the operator rather than the actual sender.

The default configuration is not affected: whatsapp.groups defaults to empty and whatsapp.dm_policy defaults to self, so a group that is not explicitly listed is dropped before dispatch.

Impacted versions: >=0.4.0 AND <0.7.2

Please refer to the article below for the most up-to-date and complete information related to this AWS Security Bulletin.

Read the article

Bulletin ID: 2026-141-AWS
Scope: AWS
Content Type: Important (requires attention)
Publication Date: 10/20/2026 13:30 PM PDT

Description:

CICD-toolchain-for-serverless-applications is an AWS sample application that provisions a CI/CD pipeline for serverless applications through the AWS Serverless Application Repository. We identified CVE-2026-112407, an incorrect privilege assignment in the IAM policy of AWS CICD-toolchain-for-serverless-applications before 1.0.11 that allows remote authenticated users with push access to the application source repository to escalate privileges and access or modify resources outside the application stack in the same AWS account via application templates deployed by the toolchain pipeline.

Impacted versions: =1.0.7 AND <=1.0.10

Please refer to the article below for the most up-to-date and complete information related to this AWS Security Bulletin.

Read the article

The third post in our semantic-layer series builds a governed customer 360 model on Amazon Redshift with AtScale, then exposes it to a Strands agent on Amazon Bedrock AgentCore through the AtScale MCP server. Identity propagation, RBAC, row security, and full query audit are enforced at the semantic layer, not the prompt.

The third post in our semantic-layer series builds a customer 360 across Amazon Redshift and Amazon Aurora with Strategy Mosaic, exposes it to a Strands agent on Amazon Bedrock AgentCore through the Mosaic MCP server, and enforces per-user row and column security end-to-end.

Build a governed semantic layer for agentic AI on AWS with the Denodo Platform and Amazon Bedrock AgentCore. It federates Amazon Aurora, Amazon Redshift, and an on-premises PostgreSQL database into a single customer 360 view an agent queries in VQL, with identity propagation and row- and column-level access control enforced at the semantic layer.

Learn how Eon replaced per-tool credentials and local MCP processes across AWS, GCP, and Azure with a single authenticated Amazon Bedrock AgentCore gateway, giving more than 100 developers and autonomous agents unified access to production investigation and cutting new-engineer onboarding from two days to two minutes.

Learn how to secure multi-agent AI systems with cryptographic identity using Keyfactor EJBCA and AWS services. This reference architecture combines signed request envelopes, user-signed delegation assertions, RFC 8693 audience-scoped tokens, and real-time OCSP revocation to prove agent identity and prevent request forgery across agent-to-agent and MCP tool calls.

Hardcoded tool wiring breaks down as an agent's catalog grows from a handful of tools to hundreds. This post walks through a runtime Discover, Instantiate, Execute loop built on AWS Agent Registry, where an orchestrator discovers MCP servers and A2A agents through semantic search at runtime instead of compiling every tool at build time.

Canadian organizations can now access OpenAI GPT-5.6, GPT-6, and GPT-6.1 models on Amazon Bedrock from the Canada (Central) and Canada West (Calgary) Regions using cross-Region inference. This post covers the models and inference profiles, how to set up permissions, and how to invoke them with the Responses, Chat Completions, and Converse APIs.

Octave built a secure, multi-tenant AI foundation on Amazon Bedrock, using Cohere Embed v4 for embeddings and Amazon Bedrock AgentCore for isolated agent execution, to power a Complaint Triage Assistant that searches 2 billion regulated documents without crossing tenant boundaries.

Learn how we fine-tuned GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod, and how to decide between supervised fine-tuning (SFT) and reinforcement learning with GRPO. SFT took reasoning-language accuracy from 22% to 100%, and GRPO added measurable gains where headroom existed.

Enterprise supply chains are data-rich but insight-poor, with the data to decide scattered across ERP, monitoring, procurement, and email. This post shows how Amazon Bedrock AgentCore exposes those systems as a governed tool layer, and how Amazon Quick turns them into one conversational experience for business users, with a deploy-it-yourself walkthrough.

The Smartsheet Model Context Protocol (MCP) Server lets AI agents read, update, and act on live work management data. This post shows two ways to connect it on AWS: Amazon Quick for business teams and Kiro for developers, so your agents can work with Smartsheet without switching tools.

The Smartsheet Model Context Protocol (MCP) Server lets AI agents read, update, and act on live data. This post shows two ways to connect it on AWS: Amazon Quick for business teams and Kiro for developers, so your agents can work with Smartsheet without switching tools.

Learn how i14 trained and served Laplace 3B, a linear attention model, on AWS Trainium2 using native PyTorch. Its recurrent state stays fixed at 29.5 MB per sequence regardless of context length, avoiding the growing key-value cache of transformer models and keeping serving memory predictable.

Learn how to build a five-step Agent Engine Optimization (AEO) scoring pipeline on AWS Lambda durable functions. Each checkpointed step suspends without incurring costs between stages and retries without re-spending on prior work, so you can score any page against five AEO signals and deploy the solution with an AWS SAM template.