Welcome back! OpenAI is rebuilding ChatGPT into an app that runs your day, and it's burying the plain chatbot to do it. Nvidia handed a Korean industrial giant the brains for its robots. And the cost of capable AI showed up twice: Microsoft found seven ways to hijack an agent, and the US is scrambling to keep its best hacking AI away from China.
In today's Generative AI Newsletter:
OpenAI's everything-app: Why is OpenAI burying the chatbot that made it famous?
Nvidia's AI robots: How does a factory robot learn to sand a panel with nobody coding the moves?
Microsoft's 7 agent hacks: How do you hijack an AI agent with one plain-English sentence?
America's hacking AI: What happens when an AI finds security holes faster than anyone can patch them?
OpenAI is rebuilding ChatGPT into a superapp that does your tasks and runs other companies' software to finish them. The Financial Times reports a redesign in the coming weeks with AI agents handling multi-step jobs and partner apps living right inside it. The model is WeChat, the Chinese app that folded messaging, payments and shopping into one screen nobody ever leaves.
The reason is money, and the timing gives it away. ChatGPT has 900 million weekly users but only 50 million pay, and its 2 million business customers already drive 40% of revenue. With a public listing weeks out, that enterprise money is the number OpenAI wants in front of investors. It's serious enough to have sidelined Sora, the video app it launched last year.
An agent that acts for you can turn a small mistake into an expensive one, and OpenAI is betting hundreds of millions of people will let it try. Microsoft promised the same kind of Copilot superapp for this summer and hasn't shipped it, so the first one out the door sets the standard.
AI agents look impressive in demos.
Production is where things get complicated.
This session with Shawhin Mosadeghzad explores what it actually takes to deploy AI agents inside enterprise workflows with governance, live data access, and human oversight.
Topics covered:
→ Why AI projects get stuck in pilot mode
→ The data foundations agents really need
→ How companies can scale agents more practically
Coursera completed its $2.5 billion all-stock merger with Udemy, combining two of online learning's biggest names into one platform built for the AI era.
Both started as long shots no VC wanted.
Coursera was two Stanford professors with no business model.
Udemy's founder heard no from more than 30 investors before building it anyway.
The returns went to whoever backed companies like these before the bet looked obvious.
That early window is where GenAI Academy sits now.
Launched two months ago.
Got 30,000 students in its first month.
Runs on the world's largest AI community, 14 million people.
NVIDIA, Google, Oracle, IBM and 300-plus partners behind it.
We’re running a crowdfunding campaign that includes the academy, plus more AI products you can find here. And now you can own a piece of it too, from $1,000, with up to 25% bonus shares before July 1.
In making an investment decision, investors must rely on their own examination of the issuer and the terms of the offering, including the merits and risks involved. Genai Works, Inc. has filed a Form C with the Securities and Exchange Commission in connection with its offering, a copy of which may be obtained here.
In Korea this week, Jensen Huang gave Doosan the full Nvidia robotics stack, the chips, the simulators and the models, to turn its factory arms into robots that figure out a job on their own.
Doosan is a Korean maker of robot arms, Bobcat diggers and power gear, and it's now building an "Agentic Robot OS" that trains robots inside a physics simulation before they touch the floor.
That's the part that matters. A robot runs a million practice reps in software where mistakes are free, then does the real thing, which is how a fiddly job like sanding works without anyone coding every move.
Nvidia's selling the whole bundle here, chips and simulator and models together, so a firm like Doosan skips building any of it. Doosan is one stop on a Korea-wide tear this week that also has Samsung, SK and Hyundai building Nvidia-powered AI factories.
The Siri overhaul Apple unveils with iOS 27 on Monday will reportedly do its thinking Microsoft just published seven new ways to hijack an AI agent, and most of them are plain English. No malware, just a sentence that reads as helpful while it steers the agent somewhere you never intended. That's what makes them dangerous. Every company scans for bad code, and a bad instruction walks right
A few of the techniques used:
Goal hijacking: slip an instruction into a doc or web page the agent reads, and it follows your order while it thinks it's still doing the user's.
Agents lying to agents: a hacked agent tells the lead agent it has clearance it doesn't, and gets waved through.
Screen attacks: bury text in a screenshot or page so a screen-reading agent obeys a command the human never sees.
Connector abuse: stand up a rogue MCP tool, the plug-in layer agents use to reach your apps, and the agent feeds it data or runs whatever it says.
Microsoft's fix is to treat every agent like a new hire, with a cryptographic ID and a permanent spot on the security team's test list. Most companies are plugging agents into real systems far faster than any of this gets built. A list like this is Microsoft saying out loud that the defenses aren't ready.
Two American AI models have gotten so good at finding security flaws that Washington now treats them as weapons. Anthropic's Mythos and OpenAI's GPT-5.5 Cyber have already turned up thousands of unknown holes in the software your phone and your bank run on. Politico reports the US has six to twelve months before China either steals that capability or builds its own.
The same model that helps you patch your own systems hands an attacker a way to find everyone else's flaws on autopilot. China has copied US models on the cheap before by training on their outputs, so Anthropic turned down its request to test Mythos and capped access at about 200 vetted organizations across 15-plus countries.
VSCO built its name on film-look photo presets, and now it's stacked with AI editing. AI Lab edits a photo straight from a text prompt, AI Remove wipes objects out of a shot and AI Upscale rescues blurry low-res images. It's free on iPhone, Android and the web, with a Pro tier for unlimited AI Lab runs.
Try this yourself:
Get VSCO from the App Store or Google Play, or open Studio in your browser at vsco.co.
Import a photo, open AI Lab and type what you want changed, like a new background or warmer light.
Use AI Remove to wipe a photobomber or a stray sign out of the shot.
Run AI Upscale on an old blurry photo to sharpen it for print or the web.
Who it's for: anyone who wants sharp edits without learning Lightroom, from creators to small-business owners shooting their own products.
Moonshot's $30B target: China's Moonshot AI, the maker of the Kimi chatbot, is in talks to raise up to $2 billion at a $30 billion valuation.
Nvidia's Korea haul: Nvidia will deploy more than 250,000 GPUs across South Korea, with Samsung, SK and Hyundai each building their own AI factories.
Anthropic opens Mythos to Europe: Anthropic agreed to give the EU's cybersecurity agency ENISA access to its Mythos model after months of pressure from Brussels.
GitHub meters Copilot: GitHub moved every Copilot plan to usage-based billing on June 1, so heavy AI coding now draws down a monthly token credit before it costs extra.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
🐝 Subscribe to our Atlas newsletter — trusted by 3M+ subscribers — to stay ahead of AI news across tech, education, and business.
📺 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.
🚀 Explore investment opportunities in the future of AI and join our community-backed growth journey.
