---
title: "Nov 7–13: Trusted access fuels a cloud breach, and more"
show: "AWS Drama"
canonical: "https://podcasts.robotpaper.ai/aws/ep/2026-11-13-week/"
date: "2026-11-13T03:36:07+00:00"
duration: "28:35"
duration_seconds: 1715
audio: "https://podcasts.robotpaper.ai/aws/ep/2026-11-13-week/2026-11-13-week-aws.mp3"
chapters: "https://podcasts.robotpaper.ai/aws/ep/2026-11-13-week/chapters.json"
transcript: "https://podcasts.robotpaper.ai/aws/ep/2026-11-13-week.txt"
feed: "https://podcasts.robotpaper.ai/aws/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Nov 7–13: Trusted access fuels a cloud breach, and more

AWS Drama · Nov 13, 2026 · 29 min

- Listen: [MP3, 27 MB](https://podcasts.robotpaper.ai/aws/ep/2026-11-13-week/2026-11-13-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/2026-11-13-week/
- Chapters (JSON): https://podcasts.robotpaper.ai/aws/ep/2026-11-13-week/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/aws/ep/2026-11-13-week.txt
- Show: [AWS Drama](https://podcasts.robotpaper.ai/aws/) (RSS: https://podcasts.robotpaper.ai/aws/feed.xml)

## Chapters

- `00:00:00` Intro
- `00:00:16` A five-dollar breakup
- `00:01:42` Ring starts roaming
- `00:03:20` Lambda keeps the meter
- `00:04:57` Shopping gatekeeper
- `00:06:30` The billion-pound plan
- `00:08:02` Water meets the cloud
- `00:09:38` The approval kept keys
- `00:11:14` Tokyo launch failures
- `00:12:59` The agents wrote it
- `00:14:43` Trusted access abused
- `00:16:32` The news
- `00:16:39` Nitro Sandboxes
- `00:17:57` Cognito token exchange
- `00:18:58` Cognito sessions
- `00:20:05` Cross-partition access
- `00:21:18` Firewall policy
- `00:22:28` AgentCore web search
- `00:23:40` AgentCore runtime
- `00:24:45` Daybreak on Bedrock
- `00:25:54` Trainium model serving
- `00:27:07` EKS device allocation
- `00:28:13` Outro

## Show notes

### The drama: the ten messiest AWS stories of the week (10)

- Scriptable GeoDNS: Leaving AWS for Bunny
- Ring cameras go mobile as Amazon debuts AI-powered security drone and robot
- Lambda keeps the meter
- Amazon’s AI shopping agent moves into text messages, taking orders and tracking prices
- Home Office cloud services contract may net AWS £1B
- ‘Water for life, not data’: Amazon data center plan draws Eastern Washington protests
- Coding Agents Copy Your Secrets to a Hidden File. Thousands Are on GitHub
- AWS Health: Amazon Elastic Compute Cloud in Tokyo: \[RESOLVED\] Elevated API issues and Launch Failures
- AWS gives developers a free local stand-in for its cloud, built from scratch by AI agents
- Trusted access abused

### The news (10)

- Introducing AWS Nitro Sandboxes on Amazon EKS, ECS and EC2 for hardware isolated pods, tasks and VMs
- Amazon Cognito now supports OAuth 2.0 token exchange
- Amazon Cognito now supports user pool session management
- AWS STS now supports cross-partition federation for AWS identities
- AWS Network Firewall launches next-generation firewall policy
- Amazon Bedrock AgentCore now offers a Web Search API
- The next generation AgentCore runtime is now the default
- Amazon Bedrock adds bedrock-runtime endpoint support for Daybreak models
- Deploy frontier AI models on AWS Trainium with SGLang-Neuron
- Amazon EKS Auto Mode and Karpenter now support Dynamic Resource Allocation

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

\[Rachel\] Hi, I'm Rachel, and this is AWS Drama, generated by RobotPaper dot A-I. We start with some drama, then we'll jump into the news. Let's count down the ten messiest AWS stories of the week.

\[Rachel\] Number ten: the cloud breakup with a four-dollar monthly saving. Developer Vincent Bernat published the details of moving his blog's delivery and geographical routing away from Amazon CloudFront and Amazon Route 53. CloudFront distributes web content, while Route 53 handles domain-name lookups and can direct traffic by location. In Bernat's account, the monthly bill went from five dollars to one after switching to Bunny and other services. His reasons included the price of geographical queries, health checks, and the key used to sign his domain records. Bunny also let him write JavaScript to choose the nearest servers. The migration came with homework. Returning two server addresses could send a reader to the more distant machine, because resolvers can shuffle their order. His compromise returns two servers when their distances are similar, or just the nearest with a shorter cache lifetime. Bunny also lacked the fine-grained, single-zone access policy he needed for certificate automation, so that part moved to the Rust-based Hickory server, with another domain hosted at deSEC. Bernat reports comparable performance so far, while saying he needs more measurements. This was his own blog's migration, with its own traffic and requirements. The lower bill arrived with a more complicated address book.

\[Rachel\] Number nine: the security camera wants to follow the floor plan. Amazon announced a new Ring drone and a wheeled Ring Rover this week, bringing its retail-device business back into the argument over cameras that move around indoor spaces. Ars Technica reports that both products combine cameras, sensors, vision models, and cloud processing to map rooms and navigate around obstacles. The two-hundred-and-fifty-dollar drone is targeting January twenty-seventh, pending authorization from the Federal Communications Commission. The thousand-dollar Rover is available for preorder, with shipping scheduled for December sixteenth. It is designed for single-floor businesses and cannot climb stairs. The privacy debate is already moving faster than either machine. Amazon says recording indicators and notifications show when cameras and microphones are active. It also says recordings will use Ring's Throw Away the Key Encryption. Ars cites the Electronic Frontier Foundation's assessment of that system as an improvement over Ring's previous default, while still falling short of end-to-end encryption. The article also notes that no facial-recognition feature has been announced for these new mobile devices. That separates this launch from the controversy around face recognition on fixed Ring cameras. The new products are still approaching their shipping dates, with the drone's regulatory approval outstanding. The promise is a camera that can investigate an alert; the privacy question follows it from room to room.

\[Rachel\] Number eight: the request finished, but the memory meter stayed awake. Easel founder Sean Whiteman published his test of AWS Lambda Web Functions this week. Lambda runs application code without requiring a team to manage the underlying servers; Web Functions support web servers and concurrent requests. Whiteman says six execution environments continued charging for memory for roughly five to seven minutes after their last request. Those measurements came from October, and the whole experiment cost about two cents at list prices. His concern was the same pattern across thousands of small applications with unpredictable traffic. The documentation says memory is billed throughout an environment's lifetime, including idle time. Whiteman found no configurable idle timeout. Forcing an exit after ten seconds cut metered time, but every intentional exit was recorded as a runtime failure. Easel chose its own shared compute, retaining traditional Lambda for overflow. The comparison also brought a benchmark argument. On Hacker News, user perpil challenged a cold-start speed claim based on only two paired trials. Whiteman acknowledged that sample size; his article also notes different hardware and a single-region test. He says busy applications can spread the idle cost across many requests. The dispute is about the economics of quiet applications, where a short burst of work can leave several minutes on the memory bill.

\[Rachel\] Number seven: Amazon's shopping agent gets a text invitation while rival agents meet the bouncer. GeekWire reports that Alexa for Shopping began an invitation-only rollout to a small group of U-S customers last Friday, with millions more expected to receive access in the following weeks. Customers can text a request or send a photo, and the agent searches for products, returns options, and completes a purchase with approval. An optional auto-buy setting can place an order when an item reaches a chosen price. It can also track packages, reorder previous purchases, and watch for price drops. The tension sits behind that convenient text thread. GeekWire says Amazon has moved to block outside shopping agents from Meta, Perplexity, Google, and OpenAI while expanding its own. This is Amazon's retail business, and the fight is over who gets to act as the shopping intermediary. Citing Digiday, the report says sponsored products can appear among the options sent by text, with those placements free to advertisers for now. Purchases the agent completes autonomously remain ad-free. Amazon says order updates, delivery notices, and deals will eventually join the conversation, and a chat can carry over to its app, website, and Alexa devices. The front door is becoming a text message. Amazon is still choosing which agents get through it.

\[Rachel\] Number six: Britain's cloud dependence comes with a much larger renewal estimate. The Register reports that a Home Office planning notice puts its next public-cloud hosting contract at about eight hundred and eighty-two million pounds before tax, and more than a billion including value-added tax. The existing contract, awarded to AWS in twenty twenty-three, was worth about four hundred and fifty million pounds. This is a planned procurement, with no new award announced. The proposed term runs from December twenty twenty-seven to December twenty twenty-nine, with a possible extension for another two years. A supplier engagement event is scheduled for December eighth. The Home Office declined to discuss the ongoing procurement, and The Register asked why the estimated value had risen so much. The awkward backdrop is the government's own concern about dependence on American cloud suppliers. The article recalls an internal Cabinet Office assessment saying vendor lock-in weakened the government's negotiating position. It also notes a newer promise to send more business to smaller suppliers, and the addition of local providers to the procurement framework. A place on that framework does not guarantee a contract. The Register's report is about the gap between those ambitions and the scale of the next deal. The billion-pound figure is an estimate on a planning notice; the supplier competition still has to happen.

\[Rachel\] Number five: Amazon brought a data-center plan, and residents brought water signs. GeekWire reports on Monday's meeting of the Walla Walla County commissioners in Washington, where residents challenged a proposed campus near the Columbia River. Citing the Walla Walla Union-Bulletin, the report says most public comments expressed opposition or concerns, including groundwater, agriculture, wildlife, and transparency. Supporters pointed to jobs and economic development. One sign said, water for life, not data. Amazon's proposal covers roughly one hundred and thirty-four acres, with three buildings and an estimated one hundred and twenty to one hundred and fifty permanent jobs. The company's fact sheet projects about fourteen million gallons of utility-supplied cooling water in the first year, plus roughly six million gallons annually for employee needs. Amazon says it is working on a system that would replace the cooling supply with nonpotable river water, and says it will cover the associated electricity costs. Many attendees wanted a one-year pause on new data-center development. Commissioner Bertha Clayton said she would consider a pause on new applications while officials draft rules, but said zoning changes now would be unlikely to affect the current phase of Amazon's project. The other two commissioners have previously opposed a moratorium. Environmental review is still underway. The protest happened; a countywide ban has not been adopted.

\[Rachel\] Number four: don't ask again became don't forget the password. This week's development is Anthropic's response to an Apiiro investigation first published on November fifth. The researchers found credentials inside public copies of Claude Code's local settings file. Their explanation is that a developer approves a shell command containing a secret and selects the option to stop asking for permission. The full command can then be stored in the project's allowlist, including that secret, and the file can end up committed to GitHub. Apiiro counted one thousand, one hundred and twenty-two current local-settings files containing credentials under its classification rules, including twenty-three AWS keys. The researchers did not use or test the credentials, so that count does not establish how many remained valid. On November seventh, Claude Code head Boris Cherny responded on Hacker News. According to the report's November eighth update, he said the team was looking into prevention and would prototype options including basic secret scanning and changes to auto mode. He also said Claude Code adds its local settings file to gitignore when it creates the file, and attributed the exposed copies to files created manually and checked in. Those are the research findings and the product team's response; a shipped fix was not announced in that exchange. A permission decision intended to save interruptions had become a place where secret-bearing commands could persist.

\[Rachel\] Number three: Tokyo's launch button said yes, and some instances still disappeared. AWS's final incident notice describes five hours and forty-five minutes of intermittent errors and connectivity problems on Wednesday, primarily in one Tokyo availability zone. Amazon EC2 is the cloud's virtual-machine service. During the first three hours of this incident, customers encountered throttling, server errors, and failed launches. Some launch requests succeeded at the API level, but the machines never started properly and moved into a terminated state. Health data for compute and block storage went missing or arrived late, sending Amazon CloudWatch monitoring alarms into insufficient-data status. Other services that depended on those launch workflows were affected too. Existing virtual machines were unaffected during that first phase. Recovery had a second act. AWS mitigated the API errors, then continued investigating network connectivity problems affecting newly launched machines and existing machines whose network configuration had changed. The notice attributes those problems to delays in an internal subsystem that propagates networking changes, related to the underlying cause. That distinction matters to the timeline: the launch A-P-Is had recovered before all the networking effects were gone. AWS reported full resolution at twelve thirty-one Pacific time on Wednesday, and said all affected services were operating normally. Its guidance was to retry failed requests and workflows. This was a confirmed regional incident, with the service's own final account documenting both phases.

\[Rachel\] Number two: AWS shipped the cloud impersonator, and says humans never reviewed its code. AWS Emulator runs simulated cloud services locally for development and testing, without an AWS account or a service fee. GeekWire reports that it draws on more than four hundred API models covering more than two hundred services. Senior principal engineer Matt Goldberg says autonomous agents wrote the implementation and continue maintaining it. One group creates tests against real AWS, another checks the emulator against those tests, and another fixes failures. Automated review gates check changes before they merge. Hacker News immediately started testing the promise behind the product. User Charlie Digital questioned whether this was an experiment or something the company would maintain. A reply from user K-E-four-Q-Q-Q said it was a real product they planned to maintain. Another commenter, scriptsmith, highlighted the documented security limits. The emulator is not intended to test access controls: its policy behavior differs from the real cloud, so local success does not establish that a request will be authorized in production. GeekWire also points to a public backlog of missing behavior and notes that the emulator does not sandbox the code it runs. AWS released the source under C-C-zero, a public-domain dedication. A local stand-in for hundreds of services now exists, while the company and its critics are still drawing the line around what a passing test proves.

\[Rachel\] And number one: the integration fell out of use, but its access was still on the guest list. Wiz published an investigation this week, carried by The Register and BleepingComputer, into attackers using credentials associated with data-platform provider Qubole to enter customer cloud accounts. The observed campaign ran from October fourteenth through October twenty-sixth. AWS Identity and Access Management lets customers create roles that trust an outside provider and specify what that provider can do. Wiz says the intruders used compromised identities in the provider's account to assume those customer-created roles. Most of the roles it observed being abused were otherwise unused; the trust remained after the integration stopped doing regular work. The attackers searched resources, collected application secrets, and downloaded data from accessible Amazon S3 storage buckets. Wiz investigated with Mandiant after Sentra detected suspicious activity. BleepingComputer reports that Sentra confirmed dozens of victims and millions of files taken from one customer. AWS's response, added to the reporting this week, says its services were not compromised and operated as designed: the attackers used third-party credentials and access customers had configured. The researchers recommend reviewing role activity, removing obsolete integrations, rotating exposed secrets, and revoking active sessions. Rotating a credential alone does not end a temporary session already issued. Qubole had not responded to the outlets when those reports were published. An unused integration had become an active access route.

\[Rachel\] And that's the drama. Those questions about access and isolation continue in this week's news.

\[Rachel\] First up: AWS Nitro Sandboxes put individual workloads inside lightweight, hardware-isolated virtual machines managed by the Nitro Hypervisor. They can host pods in Amazon Elastic Kubernetes Service, or EKS, tasks in Amazon Elastic Container Service, or ECS, and custom virtual machines on EC2. That includes agents executing model-generated code. Multiple sandboxes share the resources of one instance while retaining separate isolation boundaries. Cryptographic attestation can verify a sandbox's integrity before secrets are released to it. AWS says enabling sandbox isolation for a pod or task takes a one-line change, with existing networking, identity, storage, and monitoring integrations retained. Direct EC2 operation also supports custom machine images and orchestration, with native support for Amazon Linux and Firecracker. The launch covers most Nitro-based instances in all commercial regions, with no additional sandbox charge. The announcement describes near-native EC2 performance and support for untrusted, multi-tenant, or regulated workloads.

\[Rachel\] Amazon Cognito, the service for application sign-in and access control, now supports OAuth two point zero token exchange. A service or agent working on a user's behalf can exchange the user's token for one restricted to a downstream service. The replacement token can have narrower permissions, applies to a single audience, and comes without a refresh token. Including an actor token adds a claim identifying the intermediary, so a backend can distinguish an agent's delegated request from the user's direct request. A Model Context Protocol server, for example, can obtain a restricted token for a backend call while retaining a record of its role in the chain. Previously, intermediaries had to forward the original token or maintain their own token-minting logic. Exchange is available through the token endpoint or the new GetDerivedToken operation, on the Essentials and Plus plans wherever Cognito is available.

\[Rachel\] Cognito also added session management for user pools. Applications can list a user's active single sign-on and token sessions across devices and app clients, then revoke an individual session when needed. Each record includes the session type, app client, device type, last activity time and location, and initial sign-in time. An individual revocation can include linked sessions, while global sign-out ends all sessions associated with the user. Administrators can also cap the number of active token sessions per user. The duration for which an app client reuses a single sign-on session is now configurable, replacing the previous fixed one-hour duration. These details support custom administrative dashboards or session-management experiences inside an application. The feature is available on the Essentials and Plus plans wherever Cognito operates. Sessions can be managed through the console, command line, and A-P-Is, with settings also configurable through AWS CloudFormation.

\[Rachel\] AWS Security Token Service now supports federation between AWS partitions using temporary credentials. Partitions, such as commercial regions and GovCloud, each have an independent identity system. A workload can request a short-lived identity token in its home partition, then exchange that token for credentials to a role in another partition. The flow uses the existing GetWebIdentityToken and AssumeRoleWithWebIdentity operations. Access remains controlled by the identities allowed to assume each role, and token use can be audited in AWS CloudTrail on both sides. The announcement names consolidated security reporting and centralized operations tooling as examples. It replaces a common reliance on long-term access keys that do not expire automatically. Cross-partition federation carries no additional charge and covers commercial regions, GovCloud, China, and the European Sovereign Cloud region in Germany. Organizations already running public key infrastructure also retain AWS IAM Roles Anywhere as another cross-partition option.

\[Rachel\] AWS Network Firewall, the managed service for network access control and threat detection, has a new next-generation firewall policy model. It separates firewall access rules from intrusion prevention rules, giving each its own capacity. Access rules can enforce default-deny behavior, such as allowing connections only to approved domains. Intrusion prevention can use automatically updated managed rule groups from AWS or Marketplace partners, as well as custom groups. Rules that do not need to be shared across policies can use embedded groups, which draw capacity on demand without advance allocation. All rules in the new policy are stateful, so evaluation uses the context of the connection. The separate capacities remove the need to trade firewall rules against intrusion prevention rules within one combined allowance. Existing policies continue operating unchanged, and intrusion prevention rules remain backward compatible. The policy model is available at no additional charge in all commercial and GovCloud regions.

\[Rachel\] Amazon Bedrock AgentCore, a set of managed services for running agents, now offers a standalone web search API for applications, custom agent loops, and data pipelines. Its existing search tool ran through AgentCore Gateway. The new InvokeWebSearch operation needs no gateway or other resource created first. It returns ranked results from an Amazon-operated web index, including a text extract and, when available, the title, address, and publication time. Each request can filter by domain and publication time, including searches limited to recent announcements or particular sources. Calls use AWS identity permissions, can travel through AWS PrivateLink, and can be recorded as CloudTrail data events. Search costs seven dollars per thousand queries, matching the Gateway search rate, with no gateway charges for this path. The launch spans six regions, including Virginia and Ireland, plus four others. Applications can call the operation through the software development kits or command line, without creating a gateway search target.

\[Rachel\] New AgentCore runtimes now use the next generation by default in nineteen regions. The runtime hosts agent sessions, and this version restores each session from a snapshot of an initialized agent. Each session retains its own hardware-isolated Firecracker micro virtual machine, with capacity scaling to eight virtual processors and thirty-two gigabytes of memory. Container images can be up to ten gigabytes. AWS reports that, for a two-gigabyte image, seventy-five percent of sessions start in under two seconds, compared with twenty-nine point seven seconds on the previous generation. Idle agents scale to zero, with memory reclaimed within minutes of release. AWS says multi-turn agents can incur up to sixty-eight percent less billed memory per session. Existing runtimes stay on their current version until migrated. Setting the platform version to V2 opts into the new runtime; V1 keeps a new runtime on the previous generation.

\[Rachel\] Eligible customers can now use OpenAI's Daybreak cybersecurity models through the Bedrock runtime endpoint. Daybreak Blue, based on G-P-T five point six Sol, is intended for defensive work such as vulnerability discovery, detection engineering, and incident response. Daybreak Red, based on G-P-T five point six Cyber, supports advanced authorized research, including exploit reproduction and mitigation development. Red pairs a lower refusal threshold with stronger identity verification, monitoring, and access controls. Access requires enrollment in OpenAI's Daybreak program, followed by an AWS access request through the account team. The runtime endpoint serves the models through U-S and global cross-region inference profiles. The announcement says inference data is not used for model training, and neither model requires customers to opt in to sharing their data with OpenAI. It describes Blue as the starting point for most security teams, while Red is intended for the more specialized research tasks.

\[Rachel\] For self-hosted models, the AWS machine learning blog announced public availability of S-G-Lang Neuron, released by RadixArk with Annapurna Labs. S-G-Lang is a model-serving system; the plugin connects its existing server interface to AWS Trainium chips through native PyTorch. Applications keep the same generation A-P-Is, while the plugin handles Trainium-specific execution. It also supports separating prompt processing from token generation: one server builds the prompt's key-value cache, transfers it over Elastic Fabric Adapter, and another generates the remaining tokens. The walkthrough uses a Trainium three EC2 instance and an AWS Deep Learning Container to serve the G-P-T O-S-S one twenty B model. The plugin develops separately from the core serving framework. S-G-Lang continues managing the request lifecycle and prefix-cache bookkeeping, while the plugin supplies the hardware-specific model execution and cache-transfer path. The release includes setup instructions, supported-model documentation, and validated example configurations for running the server.

\[Rachel\] EKS Auto Mode and Karpenter now support Dynamic Resource Allocation for Nvidia graphics processors and Elastic Fabric Adapter network devices. Workloads describe the device model, architecture, or other attributes they require, and Karpenter launches the lowest-cost matching instances. The feature also supports sharing graphics processors across pods and requesting graphics and network devices with aligned P-C-I Express topology for distributed training or inference. Auto Mode includes the required device drivers. During migration, it keeps workloads using the new allocation system away from workloads using older device plugins for the same device type on a node, reducing the risk of silent oversubscription. Resource claims describe which devices may be assigned to a workload, and the workload requests those claims in its configuration. Support covers all regions except GovCloud and China. Karpenter supports all Kubernetes versions; Auto Mode requires Kubernetes one point thirty-six or newer.

\[Rachel\] Trusted integrations became an attack route, while the news brought tighter identity controls, workload isolation, and more ways to run agents. This episode was generated by RobotPaper dot A-I. I'm Rachel. Stay messy, cloud people.

---

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.
