AI Titus's Morning Wire

AI NEWS TODAY

A FREE MODEL THAT HUNTS SOFTWARE BUGS GOES PUBLIC IN TWO WEEKS
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadGoogle's cheap workhorse model got a big coding upgrade, and it is half price until New Year
Google says 3.7 Flash shows strong gains over 3.6 Flash on coding work like debugging. This is the tier most people actually run in bulk, so a discount here moves real bills.
📚 LearnNVIDIA's new small model finishes 10,000 agent tasks 30 percent faster than a model twice its size
Nemotron 3.5 Lightning is a 30B open mixture-of-experts model with 3B active parameters, hitting 86 percent accuracy while finishing 10,000 tasks 30 percent faster than Qwen3.6 35B. Licensed OpenMDW-1.1, not Apache, so read the terms before you ship on it.
★ Must-ReadOpenAI slowed down its own unreleased model because internal tests said it might be too good at hacking
The stated worry is a model that can independently find and carry out attacks on well-protected real systems. A lab hitting the brakes on itself is rare enough to notice.
★ Must-ReadOpenAI released a security-only model for approved defenders that stops refusing legitimate hacking work
Trained for finding zero-days and building exploit chains, with access gated to vetted partners. The other half of the story above.
📚 LearnIf you run Microsoft 365, there is now an admin page listing every AI agent in your tenant and flagging the ones nobody owns
A centralized view of every agent available to your organization. Worth an hour of your week, because unowned agents are the ones that quietly keep permissions after somebody leaves.
📚 LearnFree-to-download models are now only four to seven months behind the paid frontier
And once the file is on somebody's disk, the safety limits can be stripped off. That gap closing is the whole reason today's lead story matters.

📺 Watch · latest videos

📚 LearnWF2026: Software Factories & Keynotes ft. Microsoft, OpenAI, OpenClaw, Z.ai (GLM), MiniMax, HF
Featured by Latent Space.
AI Engineer
📚 LearnAgents at Scale: Inside MiniMax's Model and the Infrastructure Behind It — Dan Fu and Olive Song
Featured by Latent Space.
AI Engineer
📚 LearnFrom fork() to Fleet: Designing an Agent Sandbox Cloud — Abhishek Bhardwaj, OpenAI
Featured by Latent Space.
AI Engineer
Computer Use at the Edge of the Statistical Precipice — Pierluca D'Oro, Programma Labs
A script under one megabyte that never looks at the screen matches or beats the frontier model it was copied from. Pierluca D'Oro builds it by recordi
AI Engineer · 75 views · 2 likes
★ Must-WatchInside Cricket’s Smartest Backroom | Rajasthan Royals | ChatGPT @rajasthanroyals
What gives one of cricket’s smartest teams its edge before the first ball is bowled? Go behind the scenes with Rajasthan Royals as Kumar Sangakkara an
OpenAI · 1.6K views · 48 likes
xAI actually did it... (Grok 4.6)
Latest from Matthew Berman.
Matthew Berman · 63.8K views · 2K likes
★ Must-WatchWhat does AI actually know about you?
The information you share with AI only travels as far as you let it. Zoe from the Anthropic education team explains what happens to your information w
Claude / Anthropic · 12.1K views · 476 likes
Codex's Browser Agent Automates Literally Anything
Latest from Nate Herk.
Nate Herk · 26.5K views · 720 likes
Every Claude Code Skill I Use to Drive My Entire Development Process
Latest from Cole Medin.
Cole Medin · 10.9K views · 397 likes
Meta's new model wants "deep access" to your personal life...
Latest from Fireship.
Fireship · 407.2K views · 11.7K likes

💻 Builder Desk · Hacker News

GLM-5.3: Frontier coding with emergent cyber capabilities
791 pts · 399 comments
DeepSeek Harness developer preview
703 pts · 284 comments
Spaghettifying DRAM
677 pts · 169 comments
Understanding is the new bottleneck
398 pts · 216 comments
Choose Boring Technology (2015)
394 pts · 219 comments

🗣 Voices & Blogs

★ Must-ReadSimon Willison On Wanting Proof Before Trusting Auto Mode
He says he would love to believe the sandboxing problem is solved for Claude Code users, and then explains exactly what evidence would convince him. A good model for how to greet any vendor safety claim.
📚 LearnCharity Majors On Why AI Skepticism Stopped Being Rational
Her line is that in 2025 it was rational to be skeptical about AI, and in 2026 it is not anymore. Written for engineers who have been holding out on principle.
Zvi Mowshowitz On Pacing The Frontier Instead Of Pausing It
The case that recursive self-improvement makes every existing problem worse and adds new ones, so the question is speed rather than stop or go.
📚 LearnTom Uren On Defending A World Where Everyone Has The Model
The clearest statement of what to actually do once frontier-grade models run on any laptop. Written in June, still the best answer to today's lead.
Your container images are unsigned. In the AI era, that’s a ticking time bomb.
Most organizations that know they should sign their images still don’t. Not because they disagree, but because the path to The post Your container… · The New Stack

🌏 The Wire · Drudge / Breitbart

DEEPSEEK'S FLAGSHIP LEAVES PREVIEW, GOES LIVE FOR EVERYONE
V4-Pro is out of preview and on general release.
DEEPSEEK REWRITES ITS PRICE LIST SUNDAY, CHECK YOUR BILLS
Moving to peak and off-peak billing, with off-peak at half the peak rate. Good news if your jobs can wait.
CHIPMAKER PUSHES OPENAI'S TOP MODEL TO 750 WORDS-WORTH PER SECOND
Cerebras says no quality compromise at that speed.
APPLE BUILDS ITS OWN CHINA MODEL, FIRST FOREIGN FIRM CLEARED
Trained specifically for the China market.
CLAUDE CODE STOPS ASKING PERMISSION TODAY FOR PAID USERS
It still stops for anything irreversible, destructive, or aimed outside your environment.
Z.AI GETS TOP-TIER CODING WITHOUT TRAINING A NEW BASE MODEL
Same 743B base as GLM-5.2. Every gain came from post-training, which is a much cheaper path than most people assume.
AI DESIGNS 16 WORKING VIRUSES THAT WIPE OUT RESISTANT BACTERIA
Researchers built 285 candidates and 16 of them worked.
OPEN MODELS' GUARDRAILS COME OFF ONCE THE FILE IS ON YOUR DISK
The argument is that the policy fight is already over and the work now is preparing for it.
CHATGPT NOW REMEMBERS WHAT YOU DID ON YOUR MAC, NO SCREENSHOTS
An assistant that keeps a history of your desktop activity is a new thing to have a policy about, especially on a work machine.

🤖 Trending Models · Hugging Face

meta-models/Muse-Glimmer-30B
image-text-to-text · ★1.5K · 165.3K dl
MiniMaxAI/MiniMax-H3
image-text-to-video · ★3.9K · 2M dl
MiniMaxAI/MiniMax-Music3
text-to-audio · ★564 · 63 dl

📈 Markets

NVDA 225.30 ▲0.5%
MSFT 496.88 ▲0.9%
GOOGL 346.36 ▲0.8%
AMZN 265.13 ▼0.8%
META 594.97 ▲2.8%
AMD 483.01 ▲0.0%
AVGO 417.82 ▲0.4%
PLTR 179.01 ▲4.7%
SPCX 141.29 ▼3.3%
TSLA 339.96 ▲3.8%

The BriefTwo things happened this week and they pull in opposite directions. Models got noticeably cheaper and faster to run, and models also got noticeably better at breaking into software. Z.ai's new GLM-5.3 now tops its comparison set on CyberGym, a test of whether a model can read source code and find real, working security holes, and the company is holding the free download back about two weeks to do safety work first. The practical version for a working person: the cost of using AI keeps falling, so stop treating that as a budget question, and the cost of leaving AI out of your security plan keeps rising, so start treating that as a real one.

Level UpLearn how to tell an AI coding agent what it may never do. Coding agents are moving to act first and ask only on the scary stuff, so the useful skill this week is writing deny rules, which are a short list of commands and files the agent is blocked from touching no matter what it decides. It takes about twenty minutes and it is the cheapest safety net you can add. Configure permissions in Claude Code