The Morning Wire

AI NEWS REPORT

Latest AI news weekdays · New edition by 9 AM PT
GEMINI LANDS ON EVERY WINDOWS PC: ALT+SPACE SUMMONS IT OVER ANY APP
📚 EXPLAINER: Gemini for Windows: what a system wide AI overlay actually does on a managed PC, and the three things Google did not say
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadCursor Projects puts a coordinator agent in charge of many subagents, and it keeps working after you close the laptop
Rolling out to all users September 10. You talk to one coordinator agent and it directs the subagents. The work runs in the cloud, so closing the laptop does not stop it. Projects carry shared context files, and a project can subscribe to a Slack channel or to pull request activity. Cursor's own number: new users merged 30 percent more pull requests. No pricing stated. Beta.
★ Must-ReadA poisoned .git/config let code run outside Claude Code's sandbox, and the fix has been out since August 26
Published September 11. A repository with a bad core.fsmonitor line in .git/config made a routine git ls-files run an attacker's command outside the sandbox on macOS, with no prompt. Reported to Anthropic July 13, partly fixed August 6 in 2.1.223, fully fixed August 26 in 2.1.247. No CVE was issued. Run claude --version and make sure you are past 2.1.247.
Anthropic's September threat report says attackers now steal your AI API keys during a break-in and use them to fund the next attack
Fourth report, covering December 2025 to August 2026 across seven harm areas. The pattern that matters for anyone running networks: stolen AI API keys are being treated as a payload, not a side effect, and one breach went from first access to full admin in three hours. Anthropic's advice is to guard AI keys the way you guard production credentials, and to buy AI access only through authorized channels.
Cognition's SWE-2 coding model is post-trained from Kimi K3, a Chinese open base, two days after the US distillation advisory
Announced September 10. Built on Kimi K3, a 2.8 trillion parameter open model. Cognition reports 50.0 percent on FrontierCode 1.1 Main against Fable 5.1 at 50.9, and says it costs 64 percent less. Shipping in Devin Desktop and the Devin CLI. Weights stay closed. The Hacker News thread flags a gap worth reading: 92.8 on Terminal Bench 2.1 but 27.3 on Terminal Bench 4.
📚 LearnCohere released a translation model with open weights that it says beats DeepL and Google Translate
North Small Translate, announced September 10. 218 billion parameters total with 25 billion active, more than 50 languages, 16k context. Cohere reports 83.6 on WMT26. Weights are on Hugging Face under CC BY-NC 4.0, which means you can study it but not sell with it. Commercial use goes through RWS Language Weaver. It runs on one B200 or two H100s.
📚 LearnMistral moved 40,000 lines of Fortran 77 to C++ with more than 100 agents, and the honest part is what failed
Published September 9. A European energy company's reservoir simulator, and this was the first slice of 300,000 lines. They built a parity harness first, so the old and new code had to produce the same numbers before anything shipped. Full autonomy produced bad architecture. Putting human review gates between steps fixed it. The caveat they state themselves: the code was self-contained.

📺 Watch · latest videos

📚 LearnSAM 3: The Eyes for AI — Nikhila & Pengchuan (Meta Superintelligence), ft. Joseph Nelson (Roboflow)
Featured by Latent Space.
Latent Space
📚 LearnField Guide to Fable — Thariq Shihipar, Anthropic
Featured by Latent Space.
AI Engineer
📚 LearnAI SDK Software Factory
Featured by Latent Space.
Lars Grammel
★ Must-WatchGPT-6 Astra built a font playground
Latest from OpenAI.
OpenAI · 6.2K views · 237 likes
A Simulated Fruit Fly Is Learning to Drive
Latest from Matthew Berman.
Matthew Berman · 9.6K views · 288 likes
I Think GPT-6-Astra Just Changed YouTube
Latest from Nick Saraev.
Nick Saraev · 19.5K views · 391 likes
AI Still Needs Network Engineers
Go check out the Main video: https://youtu.be/f9IolErKgsw Sign up for Fall Into CCNA TODAY: https://ntck-ac.co/FIC
NetworkChuck · 12.6K views · 399 likes
One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer
Latest from AI Engineer.
AI Engineer · 6.8K views · 34 likes
Claude Can Now Plan Your Whole Day by Location
I gave Claude real location awareness with the MapQuest MCP. Asked it to plan a full day in NYC and it nailed the route. Go to my Instagram, DM me "AT
Nate Herk · 3.2K views · 77 likes
No One Talks Enough About Security for AI Coding. Here's How I Do It in My Workflows
AI has almost solved writing code - at least in many ways. But it certainly doesn't produce secure code! I see vulnerabilities turn up in AI generated
Cole Medin · 11.8K views · 112 likes

🗣 Voices & Blogs

★ Must-ReadHacker News on OpenAI's Agents API: is renting the agent loop convenience or lock-in?
302 points and 160 comments. One side says building your own agent harness is a huge undertaking, so paying for one is obviously right. The other side, from athrowaway3z: 'Everything else around [the API] is window dressing... I can switch providers at the drop of a hat.' And one builder, agentifysh, watching a product get absorbed: 'this was literally what i was working on for the past few months.'
📚 LearnMelanie Mitchell: stop saying 'rogue agent', because the word picks the policy
Published September 10. Mitchell argues the recent OpenAI incident was reward hacking plus missing security controls, not a model escaping. Her point is practical rather than semantic: call it an escape and you get vague rules against 'advanced AI', call it what it was and you get fixes to the engineering failures that actually caused it.
📚 LearnBen Moll and Alex Imas: AI will not hand us double digit growth, and 4 to 5 percent is the honest ceiling
Published September 9. Five reasons: adoption spreads slowly, a lot of work cannot be automated, demand and investment are weaker than the forecasts assume, security incidents cost real money, and the self-improving research loop is still unproven. Their line: 'An explosion in capabilities is unlikely to translate into double-digit growth because of the (sometimes counterintuitive) economic forces that generate GDP growth in the first place.'
📚 LearnLaura Entis: turn every correction you give a model into a test it has to pass
Published September 10. The idea is small and usable today. Every time you tell a model it got something wrong, write that correction down as a check another model can grade. Then you have a personal eval set instead of a vague sense that it is better. In her words: 'When a model falls short, you diagnose the failure, revise the skill, and rerun the eval. Failures become feedback.'
How AWS Lambda logs every flow across thousands of microVMs per host with eBPF and Rust
On any compute platform, when a security alert fires, the question is always the same. Which workload talked to that The post How AWS Lambda logs… · The New Stack

📦 What's Being Built · GitHub

affaan-m/ECC
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first dev · ★256.3K · JavaScript
NousResearch/hermes-agent
The agent that grows with you · ★244.5K · Python
deepseek-ai/deepseek-harness
DeepSeek Harness: Everything is a Plugin. · ★220.2K · TypeScript
firecrawl/firecrawl
The context API to search, scrape, and interact with the web at scale. 🔥 · ★179.1K · TypeScript

🌏 The Wire · Drudge / Breitbart

GSA SWAPS THE $1 CHATGPT DEAL FOR HALF OFF TOKENS AND NO LICENSE FEE, OPEN TO STATE, LOCAL AND TRIBAL GOVERNMENTS
Announced September 10. Starts October 1, 2026 and runs through December 31, 2028. No license fee, 50 percent off tokens, no minimum commitment. The big change is reach: it is no longer federal only, so city, county and tribal buyers can use the same pricing. Anyone serving public sector clients can quote it now.
ORACLE'S CLOUD INFRASTRUCTURE REVENUE JUMPS 121% TO $7.4 BILLION AND THE BACKLOG REACHES $664 BILLION
Reported September 10. Total revenue $19.3 billion, up 30 percent. Oracle booked more than $30 billion in new AI cloud contracts in the quarter and says it has delivered more than 300,000 GPUs since Q4. Guidance for FY27 is at least $90 billion. It also sold $20 billion of stock through an at-the-market program to fund the build.
APPLE'S A20 PRO IS THE FIRST 2NM PHONE CHIP AND ITS NEURAL ENGINE COMPUTE DOUBLES
Announced September 9. The A20 Pro in the iPhone 18 Pro is the first 2 nanometer chip in a phone, with a dual 16-core Neural Engine that Apple says doubles the AI compute of the A19 Pro. More of the model work happens on the handset instead of in someone's cloud, which is the quiet part that matters for anyone with a privacy requirement.
NVIDIA AND EIGHT AUSTRALIAN PARTNERS AIM FOR 2 GIGAWATTS OF AI CAPACITY BY 2027
Announced September 9. The partners are Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC and AirTrunk. Two gigawatts is a large amount of new capacity for one country, and it puts serious inference close to Asia Pacific customers who have been routing to the US.
EUROPE'S CYBER AGENCY IS NOW TESTING MYTHOS 5 AND GPT-6 ASTRA FOR OFFENSIVE HACKING SKILL
Reported September 10. ENISA is running the tests under Article 55 of the AI Act. It received Astra within a week of launch, while Mythos 5 took about five months to arrive. Which Mythos version ENISA actually got is not stated. This is the first time a public agency has been named as testing frontier models for attack capability.
MASSACHUSETTS TELLS BIG DATA CENTERS TO BRING CLEAN POWER OR PAY INTO A RATEPAYER FUND
Reported September 9. Any data center over 25 megawatts must supply 100 percent clean power or pay into a fund that protects other ratepayers. The state also told towns to stop signing non-disclosure agreements with developers. That is the third state in three months after Texas and New York, so this is now a pattern rather than one state's idea.
THE WHITE HOUSE FINISHED ITS FRONTIER MODEL TESTING FRAMEWORK AND IS NOT PUBLISHING IT
Reported September 10. Under the framework, companies may hand the government a model up to 30 days before release. Sam Altman says OpenAI submitted Astra. The Center for Democracy and Technology is pushing for the text to be made public. Right now nobody outside can see what is being tested or what would fail.
771 MATHEMATICIANS ASK CALTECH TO CANCEL AN AI MATHATHON BACKED BY $2 MILLION IN MODEL CREDITS
Open letter posted September 10. The event is scheduled to start October 30 and is backed by about $2 million in OpenAI and Anthropic credits. The signers say the event pushes unpublished mathematical work toward model developers. They also note the event site changed after the letter began circulating. The letter does not say any sponsor has withdrawn.

🤖 Trending Models · Hugging Face

openbmb/MiniCPM5-2B
text-generation · ★1.2K · 67.5K dl
XHToken/Spark-X2.5-4B
text-generation · ★1.1K · 17.7K dl
ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF
image-text-to-text · ★822 · 682.2K dl

📈 Markets

NVDA 218.36 ▼2.3%
MSFT 492.44 ▲0.2%
GOOGL 332.60 ▲0.6%
AMZN 251.89 ▼0.2%
META 644.38 ▼1.4%
AMD 503.60 ▼3.4%
AVGO 360.83 ▼1.0%
PLTR 165.86 ▼2.2%
SPCX 148.18 ▲0.4%
TSLA 363.56 ▼1.2%

The BriefGoogle shipped a Gemini app for Windows on Wednesday. Press Alt and Space and a Gemini window opens on top of whatever you are working on, and it can pull from your Gmail and Google Drive. Google says the app is available today globally for Windows 10 and 11, which means it can land on company laptops before anyone decides whether it should. The model runs in Google's cloud, not on the PC, so if you manage Windows machines, the thing to settle this week is whether you allow it, block it, or scope it.

Level UpThis week, open the Microsoft 365 admin center and try multi-tenant agent management. It is in public preview. It shows every AI agent across the customer tenants you manage through GDAP in one list, so you can see what is already running instead of guessing. A new AI overlay is about to be one click away on every Windows desktop you support. Knowing your current agent inventory is the cheapest way to be ready for that. →Multi-tenant agent management docs