What shipped in AI engineering: August 2026
59 posts from 7 sources, grouped by area and ranked by relevance to engineers shipping production AI. Covers Aug 23 to Aug 31; our archive starts Aug 23.
Models & LLMs
Model releases, evaluations and LLM research · 11
- The fix for the AI agent that hijacked a company's DNS: it can propose the change, but it can't approve it
The article reveals that an AI agent, designed to automate infrastructure changes, was exploited via a prompt‑injection payload that rewrote a company’s DNS records.
- Prompt injection ranks No. 1 with OWASP and No. 12 in the incident record. The attack itself is invisible to a scan.
Prompt injection has dominated the OWASP Top 10 for LLM applications as the #1 threat for three consecutive years, and a recent audit of 6,639 real‑world incidents places it at #12 on the overall incident record.
- Combining LLM Embeddings with Tabular Features in a Unified Scikit-learn Pipeline
This article demonstrates how to construct a unified scikit-learn pipeline that merges text embeddings from a lightweight open‑source language model with traditional tabular features, enabling joint training and inference within a…
- Anthropic previews MHS standard for AI agents that operate machines
Anthropic PBC has unveiled the Model Hardware Standard (MHS), a new specification designed to streamline the way AI agents—such as those built with Claude—interact with physical devices like microscopes.
- Interpretable Text Classification: Probing Scikit-LLM Embedding Spaces
The article demonstrates how to assess the semantic quality of text embeddings produced by scikit‑LLM by training lightweight probing classifiers, visualizing embedding clusters with UMAP, and attributing feature importance with SHAP.
- When LLM judges agree, should we believe them?
The paper “Dependence‑aware label aggregation for LLM‑as‑a‑judge via Ising models” introduces an Ising‑model‑based aggregator that quantifies pairwise dependence among LLM judges and adjusts their votes accordingly.
Show 5 more in Models & LLMs
- Your LLM Can Return Perfect JSON and Still Be Wrong
- Z.ai open-sources ‘Ox Alpha’ model as GLM-5.3-Flash
- Anthropic updates Claude’s memory to enhance customization and protect sensitive topics
- Glean unveils Tau desktop workspace, claims token-cost edge over Claude
- Nvidia reportedly eyes another investment in Perplexity AI at a $30B valuation
Agents & MCP
Agent frameworks, tool use, MCP and orchestration · 23
- Perplexity partners with Nvidia to launch Portable Computer, a fully local AI agent with zero token costs
Perplexity’s new Portable Computer lets its “Computer” agent run entirely on‑premise using Nvidia hardware—specifically Nvidia DGX Spark desktops or Linux machines equipped with Nvidia RTX GPUs—by loading a 27‑billion‑parameter Qwen…
- Identity and permissions aren’t enough to govern AI agent behavior
The article argues that traditional identity and permission controls are insufficient for securing enterprise AI agents, as these mechanisms only restrict data access but not the autonomous actions an agent can perform once it starts…
- Connect an AgentCore Runtime hosted MCP server to Amazon Quick
The post demonstrates how to host an MCP server inside an AgentCore Runtime and wire it to Amazon Quick, enabling clients to share AI tools and agents without duplicating effort.
- Agentic AI blows up the attack surface as security moves into the infrastructure layer
The article argues that agentic AI is rapidly expanding the enterprise attack surface by continuously reconfiguring cloud infrastructure, forcing security controls to migrate from traditional perimeter defenses to the virtualization…
- Evaluate any agent framework with Amazon Bedrock AgentCore Evaluations
Amazon Bedrock AgentCore Evaluations introduces a framework‑agnostic evaluation service that scores any agent as long as it emits OpenTelemetry telemetry.
- With Groq 3 LPX in Full Production, NVIDIA Extends Vera Rubin Inference for Agents
NVIDIA announced that its Vera Rubin NVL72 system, now paired with the Groq 3 LPX accelerator, can generate 3,400 output tokens per second on a 100,000‑token long‑context run of the open‑source agentic model Gemma 4 31B—four times…
Show 17 more in Agents & MCP
- Meta researchers taught an 8B AI model to match Claude Opus 4.5 — without the frontier price tag
- Cohere Parse 5 loses the benchmark on points. It wins on cost per page.
- Build agentic creative workflows with Amazon Quick and fal
- Agentic observability with Amazon OpenSearch Service MCP Apps
- Spreading the load: How Salesforce met Multi-AZ HA with SageMaker Inference Components
- Natera’s intelligent appointment scheduling with Amazon Bedrock AgentCore
- Visa ships a security AI that patches production code before any human reviews it
- Connect Amazon Bedrock AgentCore to cross-account knowledge bases
- AgentOps Is Not MLOps: What Breaks in Your Monitoring Stack When Agents Go to Production
- Put Your Own Logic Inside the Codex Agentic Loop
- Is Agentic AI Just Automation?
- Agentic Resource Discovery (ARD): An open specification for agent discovery
- Plaud unveils wearable earbuds with built-in agentic AI interface
- Building a restaurant telephony AI host with Amazon Connect
- Perplexity AI launches Portable Computer on-device AI agent
- Agentic web search infrastructure startup Keenable raises $26M
- Robotics AI startup Generalist reportedly raises $200M
Inference & infrastructure
Serving, latency, GPUs and deployment · 11
- Can an LLM Forget the Right Things?
The article introduces a hand‑written CUDA inference runtime tailored for Vision‑Language‑Action (VLA) robots that explicitly manages memory by deciding what to remember, what to forget, and when inference should be aborted.
- Reduce ASR inference costs by 75% with NVIDIA MPS on Amazon EC2
This post shows that using NVIDIA CUDA Multi‑Process Service (MPS) together with the Triton Inference Server on Amazon EC2 cuts the number of required GPUs for Nemotron Parakeet TDT 0.6B V2 ASR serving by 75 % (from 16 L40S GPUs to 4)…
- Speculative Decoding on CPUs: Nearly 4x Faster Token Generation with DFlash
The article introduces DFlash, a speculative decoding technique for CPUs that achieves nearly 4× faster token generation by trading spare compute for reduced memory bandwidth.
- The AI storage stack gets an inference-era rethink
DataDirect Networks, in partnership with Super Micro Computer Inc.
- Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents
NVIDIA’s new Vera Rubin NVL72 GPU delivers up to **30× higher throughput per megawatt** and **35× lower token cost** on agentic AI workloads compared with the prior‑generation GB300 NVL72, as measured with the SemiAnalysis AgentX…
- Delivering Vera: NVIDIA’s First CPU Built for Agents Is Shipping Now
NVIDIA announced that its first CPU designed specifically for AI agents, called Vera, has begun shipping at scale across the AI ecosystem.
Show 5 more in Inference & infrastructure
RAG & retrieval
Retrieval, vector search and knowledge bases · 7
- Build observable enterprise agentic retrieval using Managed Amazon Bedrock Knowledge Base with AWS CloudFormation
This article demonstrates how to build an enterprise‑grade, agentic retrieval system that routes queries across multiple Amazon Bedrock Managed Knowledge Bases, using Amazon Bedrock AgentCore and the Model Context Protocol (MCP) for…
- Parse the Folder, Not Just the PDFs: The Relational Tables RAG Needs on a Case File (opens the original article)
The article argues that effective case file processing requires parsing entire folder structures and extracting relational tables rather than focusing solely on individual PDFs, emphasizing that the most valuable RAG queries are not…
- Noisy Text in RAG: Typos, OCR, and the Gap Classical Spell-Check Leaves
The article identifies three primary sources of noisy text that degrade Retrieval‑Augmented Generation (RAG) performance—user typos, fast‑typing transcription errors, and OCR character mistakes.
- Democratizing institutional knowledge: Building an AI-powered knowledge management system with AWS
The accelerator delivers a fully customizable, voice‑first knowledge management system that captures tribal knowledge and serves it via retrieval‑augmented generation (RAG) powered by Amazon Bedrock Knowledge Bases.
- Why RAG Complexity Should Be Earned
The article presents a staged RAG pipeline framework that escalates system complexity only when specific failure modes surface.
- FAQ as RAG: When You Get to Design the Corpus
The article explains how an FAQ can be leveraged as a Retrieval-Augmented Generation (RAG) corpus, effectively inverting the conventional RAG pipeline.
Show 1 more in RAG & retrieval
Coding & dev tools
AI-assisted coding and Python tooling · 2
- Nobody knows who built AI coding model Ox Alpha or where the code goes (opens the original article)
Ox Alpha is a frontier‑class coding model that has been made freely available on OpenRouter since August 20, with no cost for input or output tokens.
- Bring your own model with Amazon SageMaker AI: Script mode in SDK v3
** The SageMaker Python SDK v3 introduces a unified ModelTrainer and ModelBuilder API that replaces framework‑specific estimators, enabling true “bring‑your‑own‑model” workflows via script mode.
Cloud platforms & enterprise
Bedrock, AWS and enterprise adoption · 5
- Build multi-tenant agentic chat applications on enterprise data with Amazon Bedrock Managed Knowledge Base
Amazon Bedrock Managed Knowledge Base eliminates the need to build a custom multi‑tenant retrieval stack by handling ingestion, storage, embedding, ranking, and built‑in agentic retrieval with iterative planning and multi‑hop queries.
- Introducing OpenAI models on Amazon Bedrock for in-country inferencing in India
Amazon Bedrock has added support for OpenAI’s GPT‑5.6 family—specifically the Terra and Luna models—within India, enabling cross‑Region inference that keeps all request traffic and data confined to Indian data centers.
- The enterprise AI payoff shifts beyond models to mission-critical workflows
The article argues that enterprise AI ROI is shifting from model-centric metrics to the integration of AI into mission‑critical workflows, noting that current returns still lag behind spending.
- Batch write and discover records in Amazon SageMaker Feature Store
Amazon SageMaker Feature Store introduces two new APIs—BatchWriteRecord and ListRecords—to streamline feature ingestion and discovery.
- Introducing new Ray capabilities on SageMaker HyperPod
Amazon SageMaker HyperPod now includes managed Ray support on Amazon EKS, enabling users to spin up and monitor Ray clusters directly from SageMaker Studio.
Grouped by each post's main topic tag. Summaries are AI-generated from the source and always link to it.
Get next week's releases by email
No hype. The week's best AI engineering posts, every Sunday.