Bedrock Brief 14 Oct 2026

Pleri AI

Pleri AI

— 4 min read
Share
Bedrock Brief 14 Oct 2026

AWS just spent the week reminding everyone that giving AI agents free rein is basically handing your toddler the car keys and hoping for the best. Between researchers at Zenity Labs and Palo Alto Networks finding ways to hijack AWS AI agents through single prompts and steal cloud credentials, AWS dropped Strands Box, an open-source sandbox designed to stop agents from going full YOLO mode. The timing feels less like coincidence and more like someone frantically patching holes while simultaneously shipping the life raft. To AWS's credit, they're moving fast on fixes, but the constant whack-a-mole with agent security flaws shows just how tricky it is to let AI do its thing without occasionally deleting your production database or racking up a five-figure API bill.

Meanwhile, AWS CFO John Felton revealed that his own pricing team ditched their nightmarish 18-tab Excel monstrosity in favor of an AI chatbot that actually makes deal modeling less painful. It's a small but telling example of where AI adoption is headed: away from "let's make the spreadsheet slightly faster" and toward "why are we even using a spreadsheet for this?" Felton says customer conversations have shifted from cost-cutting and productivity plays to building entirely new revenue streams, which sounds great until you remember those same customers are still early in their cloud migrations and probably haven't thought through the security implications of their shiny new agents.

On the robotics front, AWS launched an open-source Physical AI Toolchain with NVIDIA that stitches together data generation, model training, simulation, and deployment for robots. It's not a resurrection of the dearly departed RoboMaker, but it does pull from lessons learned wrangling over a million robots in Amazon's fulfillment centers. The big idea here is solving the synthetic data problem, because unlike large language models that can feast on the entire internet, robots need to learn from scenarios that don't exist yet. Whether this makes building AI-powered robots easier or just gives developers more places to debug their agents remains to be seen.

Fresh Cut

  • You can tell your AI coding assistant to fine-tune and deploy machine learning models on SageMaker using natural language, and it will handle the infrastructure setup automatically. Read announcement →
  • AWS EC2 G7 instances with NVIDIA RTX PRO 4500 Blackwell GPUs launched in Frankfurt and Stockholm, offering 4.6x faster AI inference than the previous generation for running language models, video analysis, and graphics rendering workloads. Read announcement →
  • Lambda Web Functions lets you run standard HTTP server code (like Express apps) that listens on a port and handles requests directly, while only charging for active CPU time instead of time spent waiting on databases or API calls. Read announcement →
  • Amazon's AI agent memory now charges by the gigabyte ($1 to ingest, $0.20 to retrieve, $0.10 monthly to store) instead of a flat fee per event, which means apps that save small chunks of data frequently will pay less. Read announcement →
  • AWS's fastest S3 storage tier that delivers single-digit millisecond response times (10x faster than regular S3) expands to London, Jakarta, Hong Kong, and Cape Town for applications like ML training that need ultra-low latency. Read announcement →
  • You can write data validation rules directly in your Spark jobs on EMR that automatically check data quality as it moves through your pipeline, catching bad records before they break things downstream. Read announcement →
  • Amazon Bedrock's OpenAI models can return a reasoning.summary parameter that explains in plain language how the model solved your coding or analytical problem, making it easier to debug your AI applications and understand what went wrong. Read announcement →
  • Claude Opus 5.5 completes coding tasks with 40% fewer tool calls than its predecessor and Claude Sonnet 5.5 runs 30% faster, both with 1M context windows, and you can use them in AWS GovCloud's Kiro IDE if you work with government data. Read announcement →
  • Amazon Connect's performance evaluation tool now validates your evaluation questions and suggests rewording when they're too vague for AI to score consistently, like specifying exactly what counts as a "sales disclosure" instead of just asking if one was provided. Read announcement →
  • Amazon Quick can extract colors, fonts, logos, and layouts from your uploaded PowerPoint or Word templates and enforce those brand rules when generating presentations and documents. Read announcement →

The Quarry

How Postman runs Agent Mode for 40 million developers on Amazon Bedrock

Postman didn't just slap an AI agent into their API platform and call it a day—they built Agent Mode on Amazon Bedrock to handle 40 million developers with real architectural discipline. The secret sauce includes ruthlessly controlling tool sprawl (because agents with too many tools make terrible decisions), exposing schema-based read operations so the agent understands API structure without guessing, and treating context window management as the actual performance bottleneck rather than an afterthought. This is a rare look at production agent patterns that matter when you're operating at scale, not just running demos on your laptop. Read blog →

More posts:

Core Sample

Can agents run research and robots?

This episode demonstrates two ambitious agent use cases: building a multi-source research agent with Amazon Bedrock AgentCore that can search arXiv, GitHub, and Hacker News while responding in different personas (noir detective, anyone?), and training humanoid robot policies using the AWS Physical AI Toolchain with NVIDIA GR00T fine-tuned in SageMaker and validated in Isaac Lab Arena. The robot workflow includes a particularly clever "success gate" that policies must pass before being registered, preventing half-baked models from making it to deployment. Both demos explore the frontier of agents that don't just chat but actually do things, from running autonomous research experiments to potentially serving as general-purpose tools for AI systems to manipulate the physical world. Watch video →

More videos:

Read more