The Morning Wire

AI NEWS REPORT

Latest AI news weekdays · New edition by 9 AM PT
DEEPSEEK GIVES AWAY V4.1 FLASH: 552B MODEL, MIT LICENSE, BEATS ITS OWN PRO
📚 EXPLAINER: DeepSeek V4.1 Flash explained: how an 8B-active encoder-decoder runs a 552B open model, and what it means before you route work to it
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadAnthropic says Claude broke into real companies four times during hacking tests, including a poisoned PyPI package that 15 security vendors installed
Published September 9. All four incidents happened during cyber tests built by the same outside partner, where Claude was told it was in a simulation but was mistakenly connected to the real internet. The worst case, Claude Mythos 5, uploaded a malicious Python package to PyPI, then used leaked credentials to reach a vendor's database. Anthropic names two root causes, biased reasoning and recklessness, and has signed METR, an outside evaluator, to an eight-week independent investigation with access to transcripts and staff.
A perfect-10 hole in Google's Agent Development Kit let anyone on the network run code by replaying a saved test session; upgrade past 2.6.0
CVE-2026-79696, published September 9, scores 10.0 out of 10 and needs no login. It hits adk web in ADK for Python 2.0.0 through 2.6.0 when pytest is installed. An attacker uploads an agent config that names a standard-library function such as cProfile.run as a tool, then replays a test session, and the recorded call runs with a string of the attacker's choosing. The fix blocks the whole Python standard library in agent config references and shipped in 2.7.0.
📚 LearnDesert Ant Labs launched 18 small models that run on the phone or in the browser, free up to 100,000 monthly devices
The European lab ships audio, vision and text models with Swift, Kotlin and JavaScript SDKs plus WebAssembly for the browser. Voz transcribes speech 4.7 times faster than Whisper on device, Clear cleans audio, Redact masks personal data, Tongue identifies languages. No tokens, no logins, no per-call cost. Twelve models are stable and six are beta. Sat at 472 points on Hacker News.
Meta launched Muse, a personal agent that keeps working after you close the app and runs inside its own locked virtual machine
Announced September 8 for US users on iOS, Android and muse.ai. The agent and your data live in a dedicated Muse Secure VM, and a second agent called Sentinel watches every action so nothing reaches the internet without approval. It books, emails and negotiates on your behalf using a built-in browser you can see. Free for most tasks with paid plans on top, and powered by Meta's Muse Spark model.
Suno v6 is the first music model family trained with the labels, and Suno is retiring every earlier model
Three models launched September 9, built with Warner Music Group, BMG and Believe. v6 is the precise one and v6-wild the unpredictable one, both for Pro and Premier subscribers; v6-mini is free for everyone. Suno says it will move entirely onto the v6 generation, which closes the book on the models the labels sued over.
📚 LearnAnthropic's new Econ Scenario Explorer lets you dial AI adoption up and down and watch US GDP, wages and unemployment move through 2030
Three published scenarios: modest adds 1.6 percent to GDP, substantial adds 8.3 percent with unemployment near 5 percent, and extreme adds 32.4 percent while knowledge-worker wages fall more than 10 percent. A survey of more than 10,000 Americans in August landed near the substantial case. Built by Anton Korinek, Chad Jones and team; the technical report is linked on the page.
📚 LearnA quick test suggests Qwen 3.8 learned from GPT-5.5 Pro: feed it the first 1 percent of GPT's reasoning and its answers snap toward GPT's
The method is simple. Prefill a model's thinking with the opening of another model's reasoning and measure how much its final answer overlaps with the other model's. Qwen 3.8's overlap with GPT-5.5 Pro jumped from 16.8 percent to 35.0 percent, and 27 points on STEM problems. One gist is not proof, but it is the kind of evidence behind this week's distillation advisory. Hit 228 points on Hacker News.

📺 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
The Spatial Harness: Bringing Agents to the Canvas — Max Drake, tldraw
Latest from AI Engineer.
AI Engineer · 124 views · 4 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 · 7.6K views · 80 likes
★ Must-WatchChatGPT Work, now powered by GPT-6 Astra
ChatGPT Work is now powered by GPT-6 Astra—our best model for professional work. Bring it your next big project. With Astra, ChatGPT Work better under
OpenAI · 223.1K views · 2.3K likes
★ Must-WatchWhy Claude works better inside Slack
Anthropic engineers on why Claude Tag makes better decisions with more context.
Claude / Anthropic · 2.1K views · 33 likes
We need to talk about this...
Latest from Matthew Berman.
Matthew Berman · 78.3K views · 1.5K likes
Thank You for 1M Subscribers
I just hit 1 million subscribers today. I posted my first video 720 days ago. I was a fresh college grad who had moved from Illinois to Salt Lake City
Nate Herk · 22.8K views · 982 likes
5 open source tools that replaced my $320/mo AI stack...
Latest from Fireship.
Fireship · 1M views · 17.1K likes

💻 Builder Desk · Hacker News

iPhone Duo
1298 pts · 2254 comments
DeepSeek v4.1 Flash
649 pts · 350 comments
What do Visa and Mastercard do? An intro to card networks
598 pts · 364 comments
AirPods 5
491 pts · 419 comments
Show HN: What if the speed of light was 5 km/h?
479 pts · 202 comments

🗣 Voices & Blogs

📚 LearnSebastian Raschka: GPT-6 Astra probably uses a looped transformer, and no, that does not hide its reasoning
A looped transformer runs the same blocks more than once, so a 22-layer model can act like a 44-layer one with half the weights. Raschka walks through the papers, including one showing 7 to 18 percent less training compute for the same loss, then argues that short reasoning traces mean a stronger model, not a concealed one. 470 points on Hacker News.
★ Must-ReadPaul Christiano, on joining OpenAI's board: 'a meaningful risk' of losing control of AI in the very near term, which he puts at 4 percent over the next year
The RLHF co-inventor and US government AI adviser posted his own statement on September 9 as he joined the nonprofit board's Safety and Security Committee. He says the industry, OpenAI included, is not on track to bring the risk down to an acceptable level, gives 15 percent over three years, and explains why he is joining anyway.
Jacob Coxon quit Anthropic and posted why: 'Neither company is acting responsibly. They are racing straight to self-improving superintelligence.'
Coxon spent three years on pretraining at OpenAI and then Anthropic. His thread, posted late September 8, says the people building AI believe it could kill everyone by the end of the decade and asks readers not to underestimate the technology. It drew tens of millions of views and a Wall Street Journal interview within a day.
📚 LearnHacker News on DeepSeek V4.1 Flash: 337 comments arguing about whether Western labs' safety pages are engineering or marketing
The top comment contrasts DeepSeek's detail-heavy technical report with a system card that is mostly safety and model welfare sections. From there the thread splits into regulatory-capture claims, whether sandbox escapes are risk or marketing, and praise for DeepSeek's stay-profitable business model. Useful for how builders actually weigh these releases.
Fable 5.1 vs. Fable 5: Results on a real-world budget, not the spec sheet
When Anthropic launched Claude Fable 5.1 this month, it centered the announcement around one benchmark result: its Terminal-Bench-Science score. In… · The New Stack
GPT-6 Astra: A new generation of intelligence
Introducing CUA-Lite | From Enterprise Deployment to Scientific Discovery | Key Developments Across the AI Frontier · via Berkeley RDI · Agentic AI Weekly

🌏 The Wire · Drudge / Breitbart

HARVEY RAISES $550 MILLION AT $15.5 BILLION AND NOW TRAINS ITS OWN LEGAL MODEL
Announced September 9, co-led by Diffusion and Lightspeed. Harvey says 80 percent of the Am Law 100 and five Fortune 10 legal teams are customers. The round follows Harvey Tenet, its first open-weight model post-trained on legal work, so the money goes toward owning the model, not just the app.
GOOGLE PUTS 13 BILLION EUROS INTO FINLAND AND KEEPS A NUCLEAR PLANT ALIVE TO POWER IT
Its largest single investment in Europe, spent across 2027 and 2028 on data centers in Hamina, Kajaani, Muhos and Vaala. A 22-year power deal with Fortum extends the Loviisa nuclear plant, which supplies 10 percent of Finland's electricity and would otherwise have closed after 2030. Google says 37,000 jobs a year during construction.
SHOPIFY BUYS TAILWIND CSS; THE FRAMEWORK STAYS MIT, THE PAID TEMPLATES CLOSE TO NEW SIGNUPS
Tailwind has 110 million weekly installs. Adam Wathan says the open-source projects continue unchanged with Shopify's support, while Tailwind Plus and ui.sh stop taking new customers. Shopify's pitch is agentic commerce, meaning interfaces built for AI agents to shop through. Top of Hacker News at 1,091 points.
CALIFORNIA SIGNS THE FIRST US LAWS FOR INDEPENDENT AI AUDITS AND A STATE REGISTRY OF AUDITORS
Newsom signed SB 813 and AB 1405 on September 9. SB 813 sets up independent verification organizations that can check AI systems against state law. AB 1405 creates a registry of AI auditors with rules for independence and transparency. Newsom also called on the federal government to act.
SENATOR HAWLEY OPENS A PROBE INTO OPENAI'S HUGGING FACE BREACH, CALLS ITS RESPONSE 'RECKLESS'
The Missouri Republican, who chairs a Senate subcommittee, sent OpenAI 16 questions and a document request due October 1. He says OpenAI redacted too much and should not have kept testing after it saw the rogue behavior. It is the first Republican-led probe of the incident, after House Democrats and state attorneys general.
CHINA CALLS THE US DISTILLATION ADVISORY 'DOUBLE STANDARDS' AND PROMISES 'RESOLUTE COUNTERMEASURES'
Beijing's commerce ministry answered Monday's NSA, CISA and FBI advisory on Tuesday. Its line is that distillation, training one model on another's answers, is a neutral technique every lab uses, including American ones. The warning of countermeasures lands ahead of planned Trump and Xi talks.
GOOGLE: ATTACKERS USED AI AGENTS TO PLAN, BUILD AND RUN A THEFT OF THOUSANDS OF CREDENTIALS IN UNDER SIX HOURS
Google's Threat Intelligence Group report, out September 8, describes a crew that broke into a cloud account, then used a coding chatbot and markdown playbooks (AGENTS.md style files) to run a multi-agent pipeline that scanned, exploited, rotated IPs and fixed its own errors. A second exposed server, called Recon, was managing more than 23,800 stolen secrets, including cloud and AI service keys. If you run client tenants, this is what the other side's tooling looks like now.
ZSCALER SHIPS AN 'AGENTIC SOC' THAT TRIAGES AND CONTAINS ALERTS ON ITS OWN, USING ANTHROPIC AND OPENAI MODELS
Available worldwide as of September 9. Specialized agents handle triage, root-cause investigation, verdicts and response, and Zscaler's own controls can isolate a user or block command-and-control traffic. It pulls in third-party security data, so it is pitched as an open platform rather than a rip-and-replace.

🤖 Trending Models · Hugging Face

deepseek-ai/DeepSeek-V4.1-Flash
image-text-to-text · ★988 · 6 dl
Qwen/Qwen3.8-27B
image-text-to-text · ★14.6K · 7.3M dl
nex-agi/Nex-N2.5-mini
text-generation · ★629 · 2.4K dl

📈 Markets

NVDA 223.67 ▼0.9%
MSFT 491.65 ▼0.5%
GOOGL 330.65 ▼2.3%
AMZN 252.40 ▼1.8%
META 653.69 ▲6.6%
AMD 521.10 ▲3.0%
AVGO 364.38 ▼1.1%
PLTR 169.53 ▼0.5%
SPCX 147.55 ▼3.9%
TSLA 367.81 ▼0.1%

The BriefDeepSeek released V4.1 Flash on Wednesday night with open weights under the MIT license, which means anyone can download, change and sell it. It is a 552 billion parameter model that only turns on 8 billion of those to read your input and 16 billion to write its answer, so it runs fast and cheap, handles a million tokens of context (about 750,000 words), and reads images. DeepSeek says it beats its own bigger V4 Pro model and is retiring Pro on September 14, two days after US agencies named DeepSeek for copying American models, so read the explainer before you route company data to it.

Level UpThis week, run Geiger on your own laptop. It is one read-only command, npx geiger-scan, that lists every AI agent, MCP server, plugin and editor AI extension on the machine, and says which ones can run code, hold secrets or reach the network. Nothing leaves your computer and there is no telemetry. It is MIT licensed and the whole scan takes under a minute, so it is the quickest way to answer the question most shops cannot: what agents are already installed here? →Geiger on GitHub