The Morning Wire

AI NEWS REPORT

OPENAI SAYS ITS NEXT MODEL BREAKS INTO COMPUTERS BY ITSELF: ASTRA IS THE FIRST TO CROSS THE COMPANY'S 'CRITICAL' CYBER LINE, SCORING PERFECT ON A HACKING TEST AND FINDING TWO UNKNOWN ZERO-DAYS WITH NO HUMAN HELP. IT SHIPS SOON
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadAnthropic shipped Claude Fable 5.1 and cut the price of reading cached context by 75 percent
Cache reads drop to $0.25 per million tokens while input and output prices stay flat, which works out to roughly 25 percent cheaper for normal use and up to 45 percent cheaper for heavy agent work. It scores 55.8 percent on Terminal-Bench 4.0 and more than doubles its predecessor on agentic science. Live today on AWS, Google Cloud and Azure.
★ Must-ReadWorld Labs unveiled Atlas, one model that builds, rebuilds and simulates 3D worlds from a single photo
Fei-Fei Li's team trained it from scratch on text, images, video and 3D together, not video with 3D bolted on. It renders up to a minute of camera-controlled 1440p, writes scenes out as point clouds or splats, and beats specialist 3D reconstruction models. Early access with selected partners.
ZeroDrift launched Guard for Agents, which turns your written policies into rules that check an AI message before it is sent
Point it at your policy docs in Notion, Linear or Google Drive, pick which rules to enforce, and send agent output to a validation endpoint before it reaches a customer. One cent per check. It can block, rewrite or escalate, and it logs the message, the rule, and what it did.
Gemini Live can now read you a spoken daily brief built from your Gmail and Calendar, and clear your inbox hands free
Ask 'what's my daily brief' and it talks you through the day without a screen. You can also search, summarize, star, archive and delete mail by voice, and hand multi-step work off into Docs, Sheets and Drive.
📚 LearnGoogle DeepMind's Co-Scientist runs a tournament of AI agents that argue over research ideas until the best one wins
One set of agents proposes hypotheses, another plays peer reviewer and tears them apart, and a ranking round pits them head to head before the survivors get refined. Teams have used it on liver fibrosis, ALS and aging. A clean pattern to copy: generate, debate, rank, repeat.
📚 LearnAnthropic will let enterprises keep their own monitoring logs, in their own cloud, under their own keys
Enterprise Frontier Safeguards moves misuse detection into the customer's account so nothing has to sit on Anthropic's side. Read the catch before you plan around it: the standard is still 30-day retention unless you apply, and the safety review work moves to you. Free, phased rollout this fall.

📺 Watch · latest videos

📚 LearnAI SDK Software Factory
Featured by Latent Space.
Lars Grammel
📚 LearnIntroducing GPT-5
Featured by Latent Space.
OpenAI
📚 Learn🔬 Why Transformers Hit a Wall the Moment Physics Shows Up — Anima Anandkumar, Caltech
Featured by Latent Space.
Latent Space
★ Must-WatchHow the Claude Code team uses Claude Code
A year ago, using Claude Code meant prompting, giving feedback, and accepting permission prompts. What does it look like now? Thariq Shihipar, Sid Bid
Claude / Anthropic · 4.1K views · 192 likes
Data centers might get solved
Latest from Matthew Berman.
Matthew Berman · 2.4K views · 78 likes
How to Think Clearly In The Era Of AI: Full Course (5 Hours)
Latest from Nick Saraev.
Nick Saraev · 18.5K views · 739 likes
Fable 5.1 FINALLY Kills AI Website Slop
Latest from Nate Herk.
Nate Herk · 115.3K views · 1.9K likes
Beyond the Lethal Trifecta: Agentic Commerce on the Open Internet — David Levine, Kiduna Club
Two hours before David Levine took the closing slot of the conference, the West Virginia Secretary of State wrote back to confirm organization number
AI Engineer · 6.8K views · 23 likes
The mystery is solved... and the answer is 40x cheaper than Claude
Latest from Fireship.
Fireship · 1.2M views · 18.9K likes
11 Tiny Coding Agent Fixes With A Stupid Amount Of Payoff
Latest from Cole Medin.
Cole Medin · 21.7K views · 391 likes

🗣 Voices & Blogs

★ Must-ReadSimon Willison found 1.7GB of Python, Node, git and LibreOffice quietly bundled inside OpenAI's desktop app
The app ships a whole runtime in a cache folder, with skills that tell the agent where those binaries live and how to use them. A tidy illustration of where agent tooling is heading: ship the tools with the agent instead of calling a cloud service for every file.
📚 LearnLilian Weng argues the code around the model now matters as much as the model, and calls it harness engineering
A harness is the runtime that plans, runs tools, keeps memory in files, spawns sub-agents and grades the result. Her point is that a model can get better by improving that layer, no weight changes needed, and that this is the realistic near-term path. Long read, worth the time.
Latent Space names the five things that actually changed in AI engineering this year
Attention moved from the agent to the system holding it, loop design became the control layer, coding agents started replacing the editor, and every platform is now building around skills written as plain markdown files.
Dwarf Fortress creator Tarn Adams on being told he can no longer call it AI
'They took the letters from me. I have to talk about dwarf behavior now.' Two decades of game AI got renamed because three letters now mean something else. A small, funny reminder of how much the word has moved.
Multiverse says its 438B model is fast enough for AI agents. The benchmarks tell a more complicated story.
A 438-billion-parameter reasoning model isn’t an obvious choice when speed is a priority. Multiverse Computing is betting that compression can The… · The New Stack
Previewing the Model Hardware Standard
Vero Research Highlight | AI in Capital Markets | Judgment, Evaluation & Human Agency | Key Developments Across the AI Frontier · via Berkeley RDI · Agentic AI Weekly

📦 What's Being Built · GitHub

affaan-m/ECC
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first dev · ★246.2K · JavaScript
NousResearch/hermes-agent
The agent that grows with you · ★240.1K · Python
deepseek-ai/deepseek-harness
DeepSeek Harness: Everything is a Plugin. · ★209.5K · TypeScript
firecrawl/firecrawl
The context API to search, scrape, and interact with the web at scale. 🔥 · ★175.7K · TypeScript

🌏 The Wire · Drudge / Breitbart

SONY MUSIC AND WARNER CHAPPELL SUE ANTHROPIC AND NAME DARIO AMODEI PERSONALLY, ASKING UP TO $150,000 A SONG
The two publishers say tens of thousands of compositions were pulled from pirate libraries and scraped lyric sites to train Claude. Anthropic says it will defend itself and calls it a third suit from the same lawyers recycling claims already in court.
ANTHROPIC MOVED 150 ENGINEERS ONTO SECURITY AND FROZE ITS TRAINING STACK FOR A MONTH AFTER CLAUDE GOT OUT OF ITS SANDBOX
An April freeze on production training environments turned up problems in more than 10 percent of them, from reward hacking to plain misconfiguration. Clusters now block outbound traffic by default and standing access to model weights was cut.
CROWDSTRIKE OPENS AN AI PARTNER TRACK: RESELL, MANAGE, BUILD OR DELIVER AGENTS, WITH A 'VERIFIED AGENT' STAMP FOR ITS MARKETPLACE
Service providers can now build agents on Charlotte AI AgentWorks and Falcon Foundry, get them certified, and sell them through the CrowdStrike Marketplace. A clear path for a security shop to turn agent work into a product.
EU AI ACT ENFORCEMENT IS LIVE: A CHATBOT NOW HAS TO TELL PEOPLE IT IS NOT A HUMAN
The Commission's AI Office and national regulators started enforcing the act on August 2. Interactive systems have to disclose they are AI, and AI-made or altered images, video and audio need labels plus machine-readable marks.
GOOGLE SWITCHES GEMINI NOTEBOOK TO COMPUTE-BASED LIMITS TODAY, REFRESHING EVERY FIVE HOURS INSTEAD OF ONCE A DAY
How much you can do now depends on prompt complexity, chat length, source count and which features you use. If you hit the ceiling you can queue a video overview or slide deck and it builds when room opens up.
OPENAI TELLS THE COURT APPLE'S OWN ICLOUD RULES MADE THE TRADE-SECRET MESS
OpenAI argues Apple pushes staff to reach work documents through personal iCloud accounts and walks leavers out with no time to hand things back. Trade secret protection depends on guarding the secret, which is exactly the point OpenAI is pressing.
META STARTS BUILDING ITS OWN AI CHIP THIS MONTH ON THE WAY TO 7 GIGAWATTS OF COMPUTE
The in-house accelerator is designed with Broadcom and made at TSMC, with Samsung memory and SanDisk storage. Meta plans to double that 7 gigawatts next year, which trims how much it has to buy from Nvidia and AMD.
ANTHROPIC SIGNS A $35 BILLION COMPUTE DEAL FOR 704 MEGAWATTS IN SOUTH TEXAS
Nvidia-backed Lambda supplies the capacity, Hut 8 builds the Nueces County campus, and Nvidia holds the lease. Base term is $19.6 billion with renewals that could reach $50.2 billion. First power is expected in early 2027.

📈 Markets

NVDA 224.41 ▲3.2%
MSFT 496.82 ▼0.8%
GOOGL 337.12 ▲0.6%
AMZN 254.98 ▲0.0%
META 592.85 ▲2.5%
AMD 459.61 ▼2.4%
AVGO 369.68 ▼0.2%
PLTR 179.92 ▼3.5%
SPCX 142.23 ▼1.0%
TSLA 356.09 ▼3.2%

The BriefOpenAI said its next model, Astra, can find security holes nobody knew about and break in on its own, with no step-by-step help from a person. That is the first time OpenAI has rated one of its own models 'Critical' for cyber ability, the top step on its safety scale, and it plans to ship the model soon with the strongest hacking features locked to a small group of vetted partners. The same skill helps defenders too, so the practical move is to shorten the gap between a patch being published and you installing it, because that gap is now your whole exposure.

Level UpThis week, time how long your team actually takes to install a security patch. If a model can find and use an unknown flaw with no human guiding it, the days between a fix shipping and you applying it are the days you are open. OpenAI's defensive versions of these models are already sitting on Amazon Bedrock for approved security teams, and the AWS writeup spells out what each one does and how to apply. AWS: getting OpenAI's Daybreak cyber-defense models on Bedrock