---
title: "Oct 17–23: Bedrock expands grounding checks, and more"
show: "AWS Drama"
canonical: "https://podcasts.robotpaper.ai/aws/ep/2026-10-23-week/"
date: "2026-10-23T06:58:03+00:00"
duration: "31:17"
duration_seconds: 1877
audio: "https://podcasts.robotpaper.ai/aws/ep/2026-10-23-week/2026-10-23-week-aws.mp3"
chapters: "https://podcasts.robotpaper.ai/aws/ep/2026-10-23-week/chapters.json"
transcript: "https://podcasts.robotpaper.ai/aws/ep/2026-10-23-week.txt"
kit: "https://podcasts.robotpaper.ai/aws/kit/2026-10-23-week/build-me-an-episode.md"
feed: "https://podcasts.robotpaper.ai/aws/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Oct 17–23: Bedrock expands grounding checks, and more

AWS Drama · Oct 23, 2026 · 31 min

- Listen: [MP3, 30 MB](https://podcasts.robotpaper.ai/aws/ep/2026-10-23-week/2026-10-23-week-aws.mp3)
- Subscribe via RSS: https://podcasts.robotpaper.ai/aws/feed.xml (paste into any podcast app)
- Web page: https://podcasts.robotpaper.ai/aws/ep/2026-10-23-week/
- Chapters (JSON): https://podcasts.robotpaper.ai/aws/ep/2026-10-23-week/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/aws/ep/2026-10-23-week.txt
- Show: [AWS Drama](https://podcasts.robotpaper.ai/aws/) (RSS: https://podcasts.robotpaper.ai/aws/feed.xml)

## Build-A-Pod (prototype)

Every story in this episode is already voiced, so an AI agent can stitch together a custom episode with no new speech. The kit is voice only. Give your agent this prompt and change the end to say what you want:

```text
Read https://podcasts.robotpaper.ai/aws/kit/2026-10-23-week/build-me-an-episode.md and build me an episode of only the AI sections
```

- Kit instructions for agents (Markdown): https://podcasts.robotpaper.ai/aws/kit/2026-10-23-week/build-me-an-episode.md
- kit.json (machine-readable kit index): https://podcasts.robotpaper.ai/aws/kit/2026-10-23-week/kit.json

Voice pieces are listed in the markdown; folders have no index page. This is a prototype: the kit format and its files may change.

## Chapters

- `00:00:00` Intro
- `00:00:11` Bedrock grounding
- `00:10:52` Serverless operations
- `00:12:11` EC2 image builds
- `00:12:43` Continuum launch
- `00:15:25` Database operations
- `00:16:36` Network connectivity
- `00:17:28` CloudWatch operations
- `00:18:30` ECS Gateway
- `00:20:26` Messaging workflows
- `00:21:24` Migration tools
- `00:22:18` Speech and stack tools
- `00:23:07` Partner and data tools
- `00:24:33` Around the web
- `00:30:34` Outro

## Show notes

### AI and machine learning (27)

- Nemotron 3.5 Super VL model now available on Amazon SageMaker JumpStart
- Grok 4.7 is now available on Amazon Bedrock in AWS GovCloud (US)
- Amazon Bedrock Guardrails now supports Standard tier for contextual grounding checks
- Amazon Bedrock Advanced Prompt Optimization now supports prompt shortening
- Structured outputs for Claude models available on Amazon Bedrock
- AWS announces TorchNeuron: Native PyTorch Support for AWS Trainium (Public Beta)
- Amazon WorkSpaces simplifies secure AI agent access to enterprise cloud desktops with less infrastructure to operate and manage
- Amazon SageMaker AI now supports custom sequence lengths for serverless model customization jobs
- Amazon Connect Customer launches two new layouts for agent workspace
- Beam model by Reflection is now available on Amazon SageMaker JumpStart and AWS Trainium
- AWS Neuron 2.33 adds native PyTorch support, Neuron Observer, and other new features
- AWS Transform adds storage migration assessments for Amazon S3, EBS, EFS, and FSx
- AWS Lambda MicroVMs is now available in 4 additional AWS regions
- Amazon SageMaker AI now supports serverless model customization for Muse Glimmer, Qwen 3.8 and Ministral 3 models
- AgentCore Gateway now supports Bedrock InvokeModel APIs and model fallback
- Amazon SageMaker AI announces training session APIs for serverless model customization
- Amazon Connect Customer now supports customizing the Customer Profiles app on the agent workspace
- Announcing new ways to customize models with SageMaker AI, without managing GPUs
- Access OpenAI GPT models on Amazon Bedrock from Canada with cross-Region inference
- bedrock-agentcore: Adds the dimension name to evaluation results and batch evaluation summaries.
- bedrock-agentcore-control: Adds model-based routing and fallback to AgentCore Gateway rules.
- devops-agent: Adds support for Microsoft Teams as a notification service.
- lexv2-models: Adds support for Assisted NLU - Dialog Manager mode allowing usage of LLMs as the default NLU system and dialog management
- polly: Added support for the new voices - Kenneth (en-US), Nate (en-US), Tina (de-DE), Carlos (es-US), Carolina (pt-BR), Leo (pt-BR), Curtis (en-AU), Kiara (en-IN),…
- bedrock-agent: Adds Retrieval Instructions for Amazon Bedrock Knowledge Bases.
- bedrock-agentcore: Adds the Builtin.FailureAnalysis evaluator for batch evaluations.
- bedrock-agent-runtime: Adds support for applying a Retrieval Instruction, optionally by version, to AgenticRetrieveStream requests for Amazon Bedrock Knowledge Bases, and reports the…

### Serverless (4)

- AWS Lambda durable functions are now available in AWS GovCloud (US) Regions
- Amazon EMR Serverless now supports IPv6 and dual-stack connectivity
- Amazon ElastiCache Serverless for Valkey now supports online migration with zero downtime
- Amazon OpenSearch Serverless now supports a 5-second default refresh interval

### EC2 (1)

- EC2 Image Builder now integrates with ASG for faster image builds

### Security Identity & Compliance (7)

- Amazon Braket now in scope for FedRAMP Class C
- AWS IAM Identity Center now supports optimistic locking for users and groups in its identity store
- Amazon GuardDuty adds optional threat detection rules for Runtime Monitoring
- AWS IAM Identity Center simplifies identity source changes from Active Directory
- AWS Network Firewall now supports Route 53 Resolved IP Validation
- Amazon Cognito extends multi-Region replication to AWS GovCloud (US) Regions
- AWS Continuum for code vulnerabilities is now generally available

### Databases (4)

- Amazon RDS for SQL Server now supports snapshots on read replicas
- Amazon RDS Custom now supports the latest Cumulative Update and GDR updates for Microsoft SQL Server
- Amazon DynamoDB global tables now support disassociating replicas
- Amazon Aurora and Amazon RDS now support On-Demand Capacity Reservations

### Networking & Content Delivery (4)

- Second-generation AWS Outposts racks now supported in the AWS Mexico (Central) Region
- AWS Direct Connect announces 100G expansion in Manchester, England
- AWS announces AWS Interconnect - last mile preview with Megaport
- AWS Direct Connect announces 100G expansion in Bangalore, India

### Management & Governance (3)

- Amazon CloudWatch log centralization now available in AWS GovCloud and additional commercial Regions
- Amazon CloudWatch Omni is now available in additional AWS Regions
- Amazon CloudWatch Alarms now supports tags in alarm actions payload

### More core services (3)

- Announcing Amazon ECS Gateway, a managed ingress solution for your ECS services
- Amazon Threat Intelligence identifies PHP2Shell exploitation prior to public disclosure in MadPot
- Amazon S3 server access log delivery to Amazon CloudWatch Logs now available in five additional AWS Regions

### Applications (4)

- AWS End User Messaging adds registration improvements for faster, higher-quality submissions
- AWS End User Messaging adds a guided wizard for RCS registration videos
- AWS End User Messaging now supports Brand Profiles
- AWS End User Messaging now offers managed OTP verification end-to-end

### Migration (3)

- AWS Transform announces Azure-to-AWS cloud-native code conversion
- AWS Transform for migrations expands support for existing landing zones
- AWS Transform now builds new networks and modernizes security groups

### Other Announcements (3)

- Polly Generative engine upgraded with more natural speech, 14 new voices, and 2 new languages
- Amazon Quick Now Available in the AWS Singapore Region
- AWS CloudFormation now simplifies stack update rollback operations

### Everything else (6)

- AWS Partner Central adds Solution Matching Engine performance insights
- AWS Partner Central launches Program Navigator, Case Study Generator, and streamlined specialization review
- AWS Clean Rooms now supports Databricks as a data source
- Amazon Connect Customer Global Resiliency now provides business users visibility into resiliency readiness across linked Regions
- AWS Service Reference Information now supports global condition keys, service names, and data plane classification
- AWS Elemental MediaConnect Router now supports high bitrate mode

### Around the web (not official AWS news; past week) (7)

- Amazon introduces toolkit for running PyTorch workloads on AWS Trainium chips
- I built an AWS agent that scores any GitHub repo before you clone it
- AWS CFO joked about the financial toll of AI. An internal recording reveals why
- Netflix Bridges AWS IAM and Internal Identity with Workload Attestation
- Google and AWS open new paths for PyTorch beyond Nvidia
- Hackers pile on PHP2Shell as AWS reveals attacks began before disclosure
- Part 2: From Docker Compose up to Terraform on AWS, and the CI That Checks It

### Music

"Sneaky Snitch" Kevin MacLeod (incompetech.com)\
Licensed under Creative Commons: By Attribution 4.0\
http://creativecommons.org/licenses/by/4.0/

The voices you're hearing are AI-generated; the stories come from public sources linked in the show notes.\
Music: Eleven Music.\
An independent project; not affiliated with or endorsed by Amazon Web Services, Y Combinator, Hacker News or any vendor or news outlet mentioned. Opinions are the hosts' own.\
Produced by Roy Osherove.

## Transcript

\[Linus\] This is RobotPaper's weekly AWS news podcast; the voices you're hearing are AI-generated; the stories come from public sources.

\[Ken\] AI and machine learning. Bedrock is a managed service for generative AI, providing access to foundation models without hosting them yourself. Guardrails checks responses against source material and the user's question. It adds a Standard tier for contextual grounding checks. It handles conversational responses and answers that reason over source material; the company also says it scores differently worded answers with equivalent meanings more consistently.

The deployment detail is cross-Region inference: Standard requires it, with a guardrail profile. Existing guardrails stay on Classic until you change them, and both tiers have the same listed pricing.

For conversational apps, this expands the response types grounding checks can evaluate beyond strict summarization, paraphrasing and question answering. The choice depends on what your application produces and whether cross-Region processing fits its deployment. Availability includes Northern Virginia, Oregon and Frankfurt, plus eight other Regions. The checks still evaluate both grounding in the supplied reference and relevance to the question. The tier also broadens language support to match the Standard content and topic filters.

\[Linus\] Bedrock Advanced Prompt Optimization tunes prompts against your evaluation criteria. It now offers optional prompt shortening, addressing cases where optimization improved a score but produced a longer prompt.

The optimizer returns a replacement only if it is shorter and scores at least as well; otherwise, you keep the original. Results show a compression ratio and both token counts.

For apps that repeat a large prompt, this provides a way to reduce input size while checking the quality measures you supplied. It supports the existing evaluation methods, including model-as-judge rubrics. Enable it on a template wherever Advanced Prompt Optimization is offered.

\[Ada\] SageMaker AI manages infrastructure for building, training and deploying machine-learning models. SageMaker JumpStart is its hub for deploying and tuning foundation models. It adds NVIDIA's Nemotron three point five Super V L, which takes images, video and text and returns text. Uses include document analysis and video questions, with context up to one million tokens.

Bedrock adds Grok four point seven in GovCloud U S for coding, agents and knowledge work. The announcement cites better handling of mixed documents and large codebases.

It also adds structured outputs for all Anthropic Claude models through its Runtime and Mantle endpoints. Apps can request JSON that matches a schema for tasks such as extracting document fields or routing support tickets, wherever those models are available.

\[Dennis\] Trainium chips run AI training and inference. TorchNeuron is a native PyTorch backend for those chips, now in public beta for Trainium two and Trainium three instances. AWS Neuron is the software used to develop and run workloads on these chips. Teams can target the Neuron device while keeping PyTorch APIs, eager execution and compilation. The backend also supports custom operators through the Neuron Kernel Interface.

For researchers experimenting with training code, the change brings familiar framework operations to Trainium without requiring a separate programming model. It remains a beta.

Release two point thirty-three includes the backend and adds the Observer health-detection tool in public beta; the company says it identifies affected nodes or devices in under a second. The release also updates inference plugins and lets Explorer trace eager-mode memory allocations back to Python source lines. Eager execution lets developers run operations directly while experimenting; compilation prepares workloads for execution on the target hardware. Custom operators remain available for hardware-specific tuning.

\[Grace\] SageMaker AI serverless model customization now accepts configurable sequence lengths, replacing fixed values in individual training recipes. Teams can set a dataset maximum for supervised fine-tuning and preference optimization, or separate prompt and response maximums for reinforcement fine-tuning.

That gives teams a direct control for matching short question-and-answer pairs or longer documents. Limits reach sixteen thousand tokens for many combinations, depending on model and technique.

It is available in Northern Virginia and Oregon, plus two other Regions. Teams can start jobs through the console or Python SDK.

Model selection also expands to Muse Glimmer, Qwen three point eight with twenty-seven billion parameters, and Ministral three in three-, eight- and fourteen-billion-parameter versions, supporting supervised, preference and reinforcement fine-tuning. Supervised fine-tuning uses labeled examples, preference optimization uses preferred answers, and reinforcement fine-tuning uses reward signals. The appropriate sequence length still depends on the published limits for the selected model and method.

\[Ken\] Reflection's Beam model is available in SageMaker JumpStart and has been ported and optimized for Trainium three. The company describes Beam as an open-weight model offered under the Apache two point zero license. Customers can inspect and modify its weights, fine-tune it with their own data, and run it inside their own cloud accounts.

Initial JumpStart availability is in Northern Virginia and Ohio. The listed checkpoint variants are B F sixteen, F P eight, and N V F P four, with deployment through a managed JumpStart endpoint. The announcement positions Beam for coding, reasoning, and multi-step agent workflows, including generating and reviewing code and automating operational tasks.

Claims about strong benchmark performance in the launch material belong to the model provider; the launch announcement does not include independent comparative results. The concrete deployment change is access to this model through SageMaker and a Trainium-optimized path. That may matter to teams that need to adapt weights and control deployment infrastructure, rather than only choose a hosted model endpoint.

\[Linus\] SageMaker AI training session APIs are generally available. Teams compose forward and backward computation, optimizer steps and sampling into a custom training loop. The model stays loaded across calls while the service handles distributed execution, checkpointing and recovery.

That lets researchers implement a custom loss, reward signal or data curriculum while leaving compute operations managed. Billing is per token processed; local reward scoring is not billed as GPU time. Availability is Northern Virginia, Oregon and Mumbai.

The company's Machine Learning Blog describes chaining techniques and agent skills for data prep, evaluation and deployment. Its training sessions cookbook supplies supervised and reinforcement-learning examples. These are recipes from the same launch for teams needing more control over training. Keeping the session state between calls matters when a training procedure needs repeated sampling and updates. The caller controls that procedure rather than choosing only a fixed training recipe.

\[Ada\] Amazon WorkSpaces provides governed cloud desktops for people and AI agents. Its managed Model Context Protocol service now handles agent sign-in and desktop links, including session-state queries and app discovery. Agents can use existing cloud credentials while the service manages each connection.

Bedrock AgentCore Gateway provides a governed endpoint for agents' model requests. It now accepts the Bedrock Invoke Model and streaming APIs for existing SDK clients.

Rules name fallback targets for throttling, errors or timeouts; a circuit breaker skips failed targets. Availability excludes China Regions.

\[Dennis\] Amazon Connect Customer provides a contact-center workspace for customer support agents. Admins can now choose Modern or Classic layouts per team: Modern moves contact controls to the top, while Classic keeps the familiar layout with resizable assistance tools. The contact-center layouts are available in Northern Virginia and Oregon, plus seven other Regions. Administrators can assign different layouts to different teams rather than changing everyone's workspace together.

Its Customer Profiles view combines contact history and customer data. Admins can build widget layouts in Profile Explorer, with calculated fields and profile summaries, and assign them to agents.

AWS Transform assesses workloads and helps prepare cloud migrations. It now accepts on-premises storage inventories, maps workloads to storage services and estimates size and cost, with what-if comparisons.

AWS Lambda runs code without customer-managed servers. Its Micro V Ms provide isolated, state-preserving compute for user or AI code. It adds Hong Kong and Malaysia, plus two other Regions, reaching fourteen.

\[Grace\] The Machine Learning Blog describes access to OpenAI G P T models from Canada Central and Canada West through Bedrock cross-Region inference. The article distinguishes U S geographic profiles, which keep processing in North America, from global profiles that can route to other supported commercial Regions.

Spotted in the SDK changelog: AgentCore evaluation results now name their dimension. Control APIs add multi-dimensional rating scales and model-based routing with fallbacks; batch failure analysis groups session failures by root cause, with suggested fixes.

AWS DevOps Agent supports operational investigation and alerts. Its APIs add Microsoft Teams team-and-channel setup.

Amazon Lex manages conversational understanding and dialog. Its Assisted Natural Language Understanding mode now uses language models for both tasks.

Amazon Polly turns text into speech. Its SDK adds generative voices across U S English, German, U S Spanish, Brazilian Portuguese, Australian English, Indian English and Korean.

Bedrock Knowledge Bases retrieves data to support model answers. Its SDK adds Retrieval Instructions with review and versions; runtime calls apply an instruction and optional version, recording both in the trace.

\[Ken\] Serverless. Lambda durable functions build multi-step workflows with checkpoints and paused execution. They now support GovCloud U S East and West, with recovery and waits that incur no compute charges for on-demand functions.

Amazon E M R Serverless runs Spark and Hive data-processing jobs without customer-managed clusters. It adds I P v six-only and dual-stack subnets, helping workloads facing I P v four address limits.

Amazon ElastiCache Serverless runs managed in-memory caches. For Valkey, it adds online migration from self-managed Valkey or Redis on virtual machines; the announcement defines zero downtime as zero client errors during migration. Migration usage is free for a hundred and eighty days after first preparing a migration, with cache charges separate. The cache migration has preparation, data synchronization and cutover stages. Applications can point to the migration endpoint first, then switch writes manually or automatically once synchronization completes.

Amazon OpenSearch Serverless provides search and vector collections without managing servers. Next-generation collections halve their default refresh interval to five seconds, making new data searchable sooner without changes to settings.

\[Linus\] Core Services, starting with compute. Amazon E C two provides virtual machines, and Image Builder creates their machine images. Image Builder can now reuse running instances from an Auto Scaling Group for builds, avoiding launch and setup time.

The group replenishes the pool as instances are used. This is opt-in: pipelines without a group keep the old launch path. The integration is available in all Regions at no added cost.

\[Ada\] Security, identity and compliance. Continuum finds and validates code flaws and proposes fixes. Its code-vulnerability service is now generally available. It builds a context graph from code and the cloud environment, including resource relationships and network reachability. For findings it can reproduce, it tests an exploit in an isolated sandbox and attaches evidence to proposed fixes.

That helps triage backlogs where a severity score may not show whether a flaw is reachable or near a critical system. Fixes can arrive as pull requests with an assessment of the change's impact.

Human approval is the confirmed starting point. The What's New entry describes a later Enforce mode, while the company's News Blog places autonomous actions in the future. The sources do not establish that autonomous fixes are available today. The context graph includes identity relationships as well as network connections. The launch describes using those relationships to rank the business impact of an exploitable finding, alongside the evidence gathered in the sandbox.

\[Dennis\] AWS I A M Identity Center manages workforce sign-in and access to cloud accounts and apps. Its identity store adds optional revision checks that reject conflicting user or group updates and deletions; membership changes are outside this check.

It also keeps identities, memberships and assignments when switching from Active Directory to another identity source. Existing identifiers remain valid in policies.

Amazon GuardDuty detects threats in workload activity. Runtime Monitoring adds more than fifty optional rules, including sensitive-file access and suspect commands, with dry-run mode for tests.

\[Grace\] AWS Network Firewall filters network traffic. Its resolved-address checks compare destinations with addresses from Amazon Route fifty-three V P C Resolver, which resolves domain names inside a virtual network, extending domain filtering to protocols such as secure shell and file transfer.

Amazon Cognito manages user and machine identities for apps. Multi-Region replication adds GovCloud, syncing identity data to a standby user pool; it requires an eligible tier and a customer-managed multi-Region encryption key.

Amazon Braket runs quantum-computing workloads. Its company-operated parts enter the FedRAMP Class C boundary in Northern Virginia, Northern California and Oregon; quantum jobs on third-party hardware remain outside that boundary.

\[Ken\] Databases. Amazon R D S manages relational databases. SQL Server read replicas now support snapshots without promotion, helping backup and recovery; point-in-time recovery from a replica is unsupported.

R D S Custom gives added control over database environments. Its SQL Server support adds cumulative and security updates for the twenty nineteen and twenty twenty-two releases.

Amazon DynamoDB is a managed key-value and document database. Its global tables provide managed, multi-Region replication. Eventually consistent tables can now detach a replica while keeping its table and data for independent use.

Amazon Aurora is a managed relational database engine. It and R D S add On-Demand Capacity Reservations for instance operations, including restores and scaling; active reservations are billed even when unused. Reservations can start immediately with no commitment and can be canceled, or be scheduled five to a hundred and twenty days ahead with a minimum commitment. Existing reserved-instance and database savings-plan discounts apply automatically.

\[Linus\] Networking and content delivery. AWS Interconnect last mile provides managed private links from customer sites to Amazon's cloud. Its preview with Megaport automates setup in San Jose and Northern Virginia. Bandwidth changes from one to a hundred gigabits per second are described for general availability, a future milestone.

AWS Direct Connect provides private physical links to the cloud. Manchester adds ten- and hundred-gigabit connections with MACsec encryption, while Bangalore adds hundred-gigabit connections.

AWS Outposts racks extend cloud infrastructure into customer data centers or colocation sites. Second-generation racks can now connect to Mexico Central, expanding their Region choices.

\[Ada\] Management and governance. Amazon CloudWatch helps monitor and troubleshoot apps using logs and other telemetry.

Log centralization copies logs across accounts and Regions into a destination account. It adds Cape Town and Hong Kong, plus thirteen other Regions including both GovCloud Regions; copies stay within each partition. The centralization announcement includes one copy of logs for free; additional copies cost five cents per gigabyte.

Omni organizes app and agent telemetry around teams and their apps. It adds Canada Central and Mumbai, plus six other Regions, reaching eleven.

Alarms can include tags in notifications so teams can route alerts without a separate lookup. The option is off by default and set per alarm; the launch excludes the United Arab Emirates and Bahrain. Alarm payloads can include all tags, selected keys, or everything except selected keys.

\[Dennis\] More Core Services, beginning with containers. E C S runs containerized apps. Gateway manages the resources that route traffic into those apps. It now provisions and manages ingress infrastructure in the customer's account. Platform teams set network standards once and link the gateway to clusters; app teams opt in.

Host-based routing lets up to twenty-five services share a load balancer, with a distinct hostname for each. The gateway adds load balancers as capacity is reached and removes them when no longer needed. Teams can inspect the resources and each service's route.

For teams with many container services, this centralizes network setup and platform standards. Support covers the task networking mode called A W S V P C, across all launch types in commercial Regions. There is no extra gateway charge, but customers pay for the infrastructure it provisions. The platform configuration covers subnets, security groups, certificates and listener settings. Services inherit that configuration when they connect, while the resulting resources remain visible inside the account.

\[Grace\] The company's Security Blog says its MadPot honeypot network recorded a PHP two Shell exploit attempt on October seventeenth, before public disclosure on October twenty-first. The actor's source of exploit details is unknown, and the report does not establish success against real systems.

Amazon S three stores objects in buckets. Server access-log delivery to CloudWatch Logs now reaches GovCloud, the European Sovereign Cloud and China, completing coverage across Regions. Teams can query access activity alongside other logs; vended-log charges apply.

\[Ken\] Applications. AWS End User Messaging manages business messages and registration. Forms now include evidence and message examples, while API responses expose conditional field requirements.

A guided Rich Communication Services wizard sends the test messages needed to record registration videos. Brand Profiles store company and use-case details for reuse across registrations, with drafts for review and optional smart matching.

Notify now generates, sends and verifies one-time passcodes over text, voice or WhatsApp. Customers set code length, expiry and attempt limits, then use separate send and verify calls, replacing their own code storage and expiry logic. The registration wizard is offered on forms that require video evidence. Brand Profiles keep the populated registration in draft form so the submitted details can still be reviewed.

\[Linus\] Migration. Transform's code conversion from Azure to Amazon's cloud is generally available for supported dot net apps. It finds Azure SDK usage and presents choices for review before converting code; Cosmos D B to DynamoDB remains in preview.

For existing landing zones, the account environments used for migrations, Transform finds gaps and builds on existing setups. An infrastructure-as-code-only mode creates plans and code, with delegated admin access for teams deploying through their own pipeline.

Network migration can build new networks or convert on-premises access rules into security groups for existing virtual networks. It finds stale rules and suggests changes without redeploying the existing network stack.

\[Ada\] Other announcements. Polly upgrades its Generative engine and adds fourteen voices across nine locales, with Hindi and Finnish now generally available. It keeps speech markup and streaming support, and expands the generative offering to GovCloud.

Amazon Quick answers workplace questions and helps users act on results. It now stores and processes data in Singapore.

AWS CloudFormation provisions and updates infrastructure from templates. Force rollback now completes a failed update rollback in one step, returning the stack to a deployable state. Resources still needing attention are recorded for review; the documentation directs customers to check configuration differences before their next deployment.

\[Dennis\] Everything else. AWS Partner Central manages partner programs and co-selling. Solution Matching Engine insights now show recommendations, attributed opportunities and stall rates, filtered by listing or type.

Program Navigator brings specialization discovery, applications and tracking together. A Case Study Generator adds assisted writing and reuse, with lighter review for general marketing stories.

AWS Clean Rooms lets partners analyze shared datasets without revealing one another's source data. It now queries Databricks tables through Delta Sharing, avoiding a preliminary export or copy.

Connect Customer Global Resiliency keeps resources synchronized across linked Regions to support switching between them. It adds a readiness page showing synchronization status and resources needing action; administrators can fix missing phone-number or user traffic-distribution links directly there.

AWS Service Reference Information supplies machine-readable metadata for policy tools. It adds global condition keys, service names and data-plane action labels.

AWS Elemental MediaConnect Router transports broadcast content. New modes support two hundred and fifty and five hundred megabits per second, up to two redundant sources and reduced jitter.

\[Grace\] Around the web and r slash AWS. These are outside reports and community work, not additional AWS announcements.

SiliconANGLE examines the work needed to move PyTorch workloads onto Trainium. Beyond the launch, its article describes tools for splitting model training across chips and built-in operators for tasks such as reducing memory use. It reports that voice startup Cartesia is using TorchNeuron for inference, while the Lieber Institute for Brain Development is exploring it for research. Those are the article's accounts of use and exploration. The piece gives teams considering the backend examples of the supporting tools and workloads to investigate.

The New Stack's article, Google and AWS open new paths for PyTorch beyond Nvidia, asks a different question: how much does getting code to run tell you about getting good performance? It compares TorchNeuron with Google's Torch T P U stack. The report distinguishes simple migration from tuning for the hardware. Its Trainium example describes changes to an attention kernel, communication and a backward pass when adapting the Laplace model to Trainium two. These are concrete places where porting work remained. The article is useful for following the gap between framework compatibility and an optimized workload; it does not establish a single performance result for every model.

\[Ken\] On dev dot to, Sarvar Nadaf presents RepoPilot, an agent built with Lambda and Bedrock that screens a public GitHub repository before someone runs it.

The concern is the installation step. The author cites research about malicious repositories disguised as useful tools, including skills and Model Context Protocol servers that coding assistants may suggest. RepoPilot scores signals from setup instructions, project history and documentation. Its safety checks look for patterns such as encoded payloads and commands that fetch and execute remote code; a critical issue limits the overall grade.

The project is worth following as an example of putting a check between an agent's suggestion and local execution. The supplied post describes what the score uses, which makes its reasoning easier to inspect. The score remains the project's assessment, not proof that the repository is safe.

\[Linus\] A Business Insider report discussed on Hacker News examines the cash-flow impact of Amazon's AI infrastructure spending. The report says Amazon is spending two hundred and twenty billion dollars this year on the physical infrastructure supporting AI.

The report attributes to infrastructure executive Prasad Kalyanaraman an expectation that the company's free cash flow will remain negative for the next couple of years. That is an attributed expectation from the reported internal meeting, not an independently verified forecast or a guarantee about future financial results.

A commenter in the Hacker News discussion argued that replacing knowledge workers could weaken demand elsewhere in the economy and complicate the industry's route to profitability. That is a commenter's economic argument, not evidence establishing the financial outcome of Amazon's investment. The report describes the scale and timing of spending, while the eventual returns remain uncertain.

\[Ada\] AWS Identity and Access Management, or I A M, controls access to cloud resources through identities and permissions. InfoQ's article on Netflix's workload attestation describes how Spark jobs on E M R obtain short-lived internal credentials without developers injecting them by hand.

Netflix's control plane signs metadata identifying the intended workload. The job also proves its cloud role to Netflix's identity service. Together, those checks link the cloud execution identity to the internal identity the job should receive.

Each Data Project has a dedicated role. With the number of projects expected to reach the tens of thousands, Netflix spreads those roles across a pool of cloud accounts. Project ownership becomes part of the identity design.

The design deals with scale at the Spark driver, the process coordinating the job. The driver establishes its identity once and distributes credentials to executors over authenticated, encrypted connections. That avoids a separate attestation for every executor, while making the driver part of the trust boundary. It also repeats attestation before the certificate expires.

The case study covers identity checks and credential distribution for teams connecting managed compute to internal services.

\[Dennis\] The Register also reports on PHP two Shell attacks before public disclosure, in its headline Hackers pile on PHP two Shell as AWS reveals attacks began before disclosure.

\[Grace\] On dev dot to, Nnamdi Felix Ibe's article, From Docker Compose up to Terraform on AWS, and the CI That Checks It, examines what a working local demo fails to prove.

The author describes a health-data conversion application and a planned cloud deployment. When checking the pipeline, he found a secret-scanning configuration with no active rules. A test with a planted credential showed that the empty configuration passed while the default rules caught it. After fixing the configuration, a scan of repository history found a previously committed credential. The author says it has been revoked.

The practical concern is whether a passing check has tested the property people think it has. The article offers a concrete investigation to follow, including the difference between scanning current files and their history.

The author also draws a deployment boundary: the version readers can clone has been planned, not deployed. A successful plan is not evidence of a running system.

\[Linus\] That concludes RobotPaper's weekly AWS news podcast. The recurring themes were more control over model customization, clearer operational context, and managed workflows for infrastructure changes. The source distinctions remain important: a public beta differs from general availability, and an announced future capability differs from behavior available now. The full write-up and source references are at robotpaper dot A I. From Linus, Dennis, Ken, Ada, and Grace, thank you for listening.

---

The voices you're hearing are AI-generated; the stories come from public sources linked in the show notes. An independent project; not affiliated with or endorsed by Amazon Web Services, Y Combinator, Hacker News or any vendor or news outlet mentioned. Opinions are the hosts' own. Produced by Roy Osherove. Privacy: We count visits to these pages with Cloudflare Web Analytics (cookieless). Google Analytics (which sets cookies): visitors in the EU/EEA, the UK and Switzerland are asked first and it runs only if they accept (we go by the country Cloudflare reports for the connection; if it is unknown, we ask). Elsewhere it loads by default. Anyone can turn it off with the Privacy settings button at the bottom of every page; the choice is kept for a year in this browser (local storage "analytics-consent"). We count feed fetches and downloads in aggregate; IP addresses are hashed and never stored. If you send feedback, we keep what you write (and your email if you give it) to read and reply. © 2026 Roy Osherove.
