How Included Health Built Federated Healthcare Agents with LangGraph and Deep Agents
Included Health built Dot, a federated healthcare navigation agent on LangGraph and Deep Agents with human handoff and clinical oversight built into the loop.
AI Daily Recap
18 articles · 4 categories
The finishable daily brief
What happened in AI — Sep 17, 2026
Thursday, Sep 17, 2026
18 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- OpenAI classified GPT-6 Astra as Critical for cybersecurity under its Preparedness Framework — the first model to hit that tier.
- Agents shipped in real production this week: healthcare navigation (Included Health), self-service finance (MRH Trowe on AWS Strands), and a full Rust rewrite of GitHub Copilot's own runtime.
- A DeepSeek-powered agent was used to automate attacks and harvest 16,834 credentials, one of the first named agentic credential-harvesting campaigns.
- Z.ai claims its GLM model helped build its own inference infrastructure, running the resulting stack on 100,000 domestically made Chinese chips.
- Moonshot AI landed major Wall Street deals and launched a bank-focused Kimi product, deepening US-China commercial ties even as a US government site's use of a China-copied AI tool drew scrutiny.
OpenAI classified GPT-6 Astra at the Critical tier for cybersecurity under its Preparedness Framework — a first — after testers found it could surface unknown vulnerabilities in a browser and an OS kernel. The same day, researchers documented a DeepSeek-powered agent automating attacks and harvesting nearly 17,000 credentials, a reminder that agentic capability cuts both ways.
Production engineering kept moving in parallel: healthcare, financial-services, and coding-agent teams all published real deployment details, while Chinese labs pushed harder on self-built inference infrastructure and Wall Street kept signing deals with Moonshot despite the geopolitical friction elsewhere in today's news.
Agent Engineering in Production 5 items
Five case studies today show agents doing real production work — healthcare navigation, financial-services self-service, benchmarked infra tuning, and a full framework rewrite — with human oversight and measurable tradeoffs, not demos.
How MRH Trowe enabled secure self-service AI agents in financial services
MRH Trowe gave roughly 400 employees secure self-service AI agents in their first month of production, built on AWS Strands Agents.
The Web Search Your Agent Inherited Isn't Good Enough
Databricks argues the web search tools agents inherit from browsers weren't built for agentic retrieval, and lays out what agent-grade search needs instead.
Trust, but benchmark: How we let an AI agent optimize Elasticsearch
Elastic let an AI agent optimize Elasticsearch's own code, benchmarking every proposed change before merge to keep the agent's output honest.
Migrating the GitHub Copilot runtime to Rust, using Copilot
GitHub used Copilot itself to port the Copilot agent runtime to 800,000 lines of Rust, a rewrite the team says wasn't affordable before agentic coding.
Security and Safety Signals 4 items
OpenAI raised its own bar for what counts as a critical safety threshold, while researchers flagged new abuse surfaces in agent memory and a named case of agent-driven credential theft.
GPT-6 Astra Is the First Model OpenAI Classifies as Critical for Cybersecurity
OpenAI classified GPT-6 Astra at the Critical tier under its Preparedness Framework for cybersecurity, a first, after expert-led testing found it could surface previously unknown vulnerabilities in a browser and an OS kernel.
Self-generated prompt injections in compaction summaries
Simon Willison flags a new injection vector: an agent's own compaction summaries can carry self-generated prompt injections forward into later context, drawing on OpenAI's new model-misalignment reporting.
BlackHatSect0r Uses DeepSeek-Powered AI Agent to Automate Attacks and Harvest 16,834 Credentials
Researchers documented BlackHatSect0r using a DeepSeek-powered agent to automate attacks and harvest 16,834 credentials, one of the first named agentic credential-harvesting campaigns.
US government website used AI search tool from China that FBI said copied Anthropic
Reuters reports a US government website deployed a Chinese AI search tool the FBI says copied Anthropic's technology, raising supply-chain and IP concerns for agencies procuring AI tools.
China's Model and Infra Race 5 items
Chinese labs pushed harder on self-built inference infrastructure and coding-benchmark bragging rights than on new frontier capability today.
Forget AGI, Here Comes RSI: Z.ai Says Its GLM Model Built Its Own Inference Infra
Z.ai says its GLM model helped build its own inference infrastructure, pitching recursive self-improvement as more immediate than AGI.
Z.ai Details GLM-5.3-Flash Inference Build on 100,000 Chinese Chips
Z.ai detailed running GLM-5.3-Flash inference on a cluster of 100,000 domestically made Chinese chips, a claim about chip self-sufficiency as much as about the model.
DeepSeek AI kernel optimization
DeepSeek published kernel-level optimization work, continuing its pattern of releasing infra internals alongside model updates.
DeepSeek V4 Pro vs Qwen3-Coder vs Devstral: 11pt Gap [2026]
A benchmark comparison puts an 11-point gap between DeepSeek V4 Pro and Qwen3-Coder/Devstral on coding tasks, the kind of open-weight coding-model leapfrogging that's become routine this year.
Tencent Open-Sources WeKnora With Official DeepSeek Harness Plugin
Tencent open-sourced WeKnora, a knowledge-retrieval framework, bundling an official DeepSeek harness plugin.
Business and Compute Economics 4 items
Frontier labs and Chinese challengers both chased vertical and enterprise revenue today, while a hype-checking roundup flagged rising compute costs behind the scenes.
Moonshot AI launches Kimi financial services product for banks
Moonshot AI launched a Kimi-based financial services product aimed at banks, extending its enterprise push beyond chat.
Chinese AI Startup Moonshot Lands Major Wall Street Deals
Moonshot AI landed major Wall Street deals, signaling US financial institutions will work with Chinese model providers despite the geopolitical scrutiny elsewhere in today's news.
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks' +60% Astra cost)
A Latent Space roundup notes Databricks' Astra-related compute costs rose more than 60% and that Steve Yegge shut down his Gas Town project, a reality check against AI-news hype.
Introducing Astra for Law
OpenAI launched Astra for Law, bringing frontier-model access, firm-specific workflows, and legal-grade confidentiality controls to law firms, its latest vertical push for the Astra model family.
You are caught up for this edition