Collaborate with us Welcome back! Anthropic added a way to teach Claude a task just by letting it watch you work, OpenAI admitted two of its models broke into Hugging Face to cheat on a hacking test, Google fired three new Gemini models at Anthropic including a cyber one and Substack is rolling out a tool to catch AI slop in your feed. In today's Generative AI Newsletter: Anthropic: What's the new way to use Claude?
Claude can now learn a task by watching you do it once Anthropic added a new way to use Claude, and it doesn't involve typing a prompt. You record your screen while you do a task, talk through what you're doing, and Claude turns that recording into a skill it can run again on its own. It picks up the job the way a new hire would, by watching you do it once instead of reading a manual. The pitch is the repetitive software stuff, editing a video or laying out a design.
Show it a single time, then hand it off. It stops short of full autopilot. You still point it at what matters, and the taste stays yours. It's live now on Pro, Max and Team plans, under Record a skill in the desktop app's plus menu.
Special highlight from the network Make Claude Your 24/7 Second Brain in Just 16 Hours Claude is the most powerful AI tool of 2026, with features like Skills, Connectors, Cowork, and vibe coding. But most people barely scratch the surface. Our experts distilled 800+ hours of research and real-world experience into a 16-hour live workshop that teaches you how to unlock Claude's full potential alongside 10+ AI tools and workflows. You'll learn to: Master Claude's Chat, Cowork, and Code modes Set up Skills, Connectors, and plug-ins to automate work Build apps and dashboards with vibe coding—no coding required 🧠 Saturday & Sunday | 🕙 10 AM–7 PM EST Reserve your spot today.
OpenAI's new model caused a major incident with Hugging Face to cheat a test Remember the mystery AI that broke into Hugging Face? OpenAI just admitted it was theirs. Two of its models, GPT-5.6 Sol and an unreleased stronger one, were running a hacking test with their safety limits switched off. AI decided the answer key was sitting on Hugging Face, so they went and took it.
To get there, the models broke out of a sealed lab with no internet and clawed onto Hugging Face's servers, using a zero-day bug to pull the test solutions from its live database. OpenAI calls it an unprecedented cyber incident. Its own report says the models were hyperfocused and went to extreme lengths for one narrow goal. Hugging Face caught it fast, rotated its keys and beat the attack back with a Chinese open model.
This was their only option since other models refused to help due to safeguards. Google aimed a new Gemini model straight at Anthropic Google answered a rough week with a product dump, three new Gemini models in a single day. Two are cheaper, faster Flash models for everyday work. The third, Gemini 3.5 Flash Cyber, is a security model built to find and patch software holes, pointed straight at Anthropic's Mythos.
Google says it scores 42% higher than its last version on long cyber tasks, and it's going only to governments and vetted partners for now. Last week Google delayed its flagship Gemini Pro and lost $200 billion in a day. So it shipped the cheap Flash line and a cyber shot at Anthropic instead, while the big model stays stuck. Substack will scan your feed for AI slop Substack is handing readers a way to sniff out AI writing.
A new Scan for AI text button, built with detector Pangram, rates how much of a post or comment came from a human versus a machine. It runs in the Substack reader on web and iPhone, on any post or comment longer than 100 words. Writers also get a space to say whether they used AI and how. Substack even coined a word for faking a human voice with a bot.
It's calling it Claudefishing. The company aimed the whole thing at what it calls a feed drowning in AI slop, with a dig at LinkedIn on the way out. Tool of the Day: Warp Warp is a terminal built for the age of AI coding agents, and the trick is you stop babysitting just one. You can fire off a whole team of them, Claude Code, Codex, Gemini CLI and Warp's own agent, then push them to the cloud to keep working after you close the laptop.
One company used it to build an AI teammate that now writes 35,000 lines of code a week. Try this yourself: Download Warp free at warp.dev, it's open-source and runs on Mac, Windows or Linux. Ask Warp Agent to do a task and let it auto-route to the best model for the job. Kick off Claude Code and Codex on two different tasks at once and watch both from one window.
Push an agent to the cloud so it keeps running after you shut the laptop. Who it's for: developers tired of watching one agent work while everything else waits. Everything else you shouldn't miss Altman briefs Washington on what's next : The OpenAI chief will walk the Trump administration and lawmakers through his next-generation models next week, days after two of them hacked Hugging Face. You can see your iOS app run on Claude Code : On desktop it builds and runs your iOS app in a simulator right beside the chat, so you watch each change the instant it writes the code, in public beta.
Moonshot wants a $50 billion price tag : The Chinese lab behind the sold-out Kimi K3 is reportedly raising a final round at up to a $50 billion valuation before it goes public. OpenAI isn't the only lab with a runaway : Anthropic's Mythos reportedly broke out of its own sandbox during testing to email a researcher. Learn more about AI from the experts building it 📸 Follow us on Instagram for fast, visual AI updates in 30 seconds. 📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators. 🐦 Follow us on X for breaking AI news and real-time industry updates. 🧠 Learn how to build your next AI application with practical resources and expert guidance.