
Welcome back! US government investigators got ChatGPT, Claude, Gemini and Grok to plan attacks in under 12 hours. Anthropic’s shares start trading today after a record-breaking sale. OpenAI says its AI made AMD’s best chip design up to 14% better. And Meta’s own staff warned that Muse was reading their private messages.
In today's Generative AI Newsletter:
Chatbots: How did investigators get past the safety rules?
Anthropic: How much is Claude’s maker worth now?
OpenAI: How long did a year of chip work take?
Muse: Whose bank balance did it share?
Common tricks beat every chatbot’s safety rules

US government investigators went undercover and got ChatGPT, Claude, Gemini and Grok to plan attacks, each in under 12 hours.
And every chatbot pushed back at first, but common tricks got them past their safety rules.
The test was run this spring by the Government Accountability Office (GAO), the independent office that investigates for the US Congress. Its report came out Wednesday.
Two chatbots planned car bomb attacks on a hotel and a convention center in major US cities, with an exact date and time.
The other two planned attacks on power stations, one with a drone and one with guns.
The investigators checked with experts and confirmed the plans could work.
The report labels the chatbots A to D, so it doesn’t say which one did what.
The companies told the GAO they will keep making their chatbots safer.
Everything the chatbots used was already out there. What used to take weeks of digging, they put together in hours.
Courses built by AI experts for $19.90 a month

For $19.90 a month, you get access to courses made by AI experts.
Every course is short and focused on tools you can start using the same day. Learners give them 4.6 out of 5.
Your subscription includes:
How to Stop Hitting Your Claude Limit: learn why Claude stops you in the middle of a task, and how to prevent it
Claude Cowork: Build Your AI Business Team: set up your first 3 AI agents to take work off your hands.
Get Your Business Out of Your Head with Claude: talk through a process for 5 minutes and turn it into a step-by-step SOP.
Ship a Micro-SaaS in a Weekend: build a small app with Stripe payments and start charging real money for it.
The curriculum is rebuilt every quarter, so your skills keep pace with the latest models.
Anthropic just beat Elon Musk’s record by $6 billion

Anthropic, the company behind Claude, sold 980 million of its shares at $83 each before they reach the stock market. The sale brought in $81.3 billion.
And no company has ever raised more this way, not even Elon Musk’s SpaceX with $75 billion.
Until now, only big investors and staff could own a piece of Anthropic. From today, anyone with a trading account can buy and sell its shares on the Nasdaq, the US stock market, under the name “ANT.”
So many investors wanted in that Anthropic raised the price above the $70 to $78 it first planned.
It also sold more shares than planned: 980 million instead of 825 million.
At $83 a share, the whole company is worth about $1.5 trillion, up from $965 billion in May.
Amazon owns nearly 15% of Anthropic, and Google almost 14%.
Each of its seven co-founders, including CEO Dario Amodei, sold nearly 3.4 million shares, worth about $282 million each.
The first trades happen later today.
Now the whole world gets to decide what Claude is worth.
OpenAI did a year of AMD’s chip work in 5 months

OpenAI says its AI models did “roughly a year of progress” on AMD’s top GPU design in less than five months.
And it made the design up to 14% better.
GPUs are the chips that run AI, and AMD’s engineers had already spent years fine-tuning these designs. OpenAI’s models went through the code that describes the chip’s circuits, rewrote parts of it, and checked every change with chip-design software from Synopsys.
The gains are in power use, speed and chip size, the three things every chipmaker fights over.
AMD has already put some of the changes into a chip it sent for manufacturing.
AMD also confirmed smaller gains, in the single digits, on its next top processor (CPU).
OpenAI used the same method on Jalapeño, its first custom chip, shrinking two of its main parts by 10% and 13%.
In OpenAI’s chip-design tests, its newest model, GPT-6.2 Astra, scored the highest.
OpenAI is now working with Synopsys to build these AI skills into the tools chip engineers already use.
AI is starting to design the chips that run AI.
How long before most chips are designed by AI?
Meta’s Muse reads chats you never shared

Meta employees have been warning each other in internal posts that Muse, Meta’s AI agent, was reading their private messages and mixing up people’s accounts.
One employee was getting fishing weather updates from Muse when it added, out of nowhere: “BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57.” Neither account was his.
Another employee got a reminder from Muse to reply to her sister on Signal, a private messaging app that competes with Meta’s WhatsApp. She had never given Muse access to it.
One said Muse booked a non-refundable Airbnb stay without asking.
Reuters tested this itself. On Android phones and Macs, Muse could read people’s Signal messages. On iPhones, it couldn’t.
Apple updated its Mac software this week to block it.
Meta calls these “isolated incidents” and says many were the AI making things up. Reuters couldn’t tell whether the bank balances were real.
Muse has been downloaded more than 10 million times since September.
Tool of the Day: Playwright

Playwright is a free, open-source tool from Microsoft that lets code, or an AI agent, use a web browser like a person: open pages, click buttons, fill in forms and check that everything works. Developers use it to test websites in Chrome, Firefox and Safari’s engine at once. Its MCP add-on lets AI tools like Claude and GitHub Copilot control a real browser for you. It works on Windows, Mac and Linux.
Try this yourself:
Go to playwright.dev and follow the Get started guide.
Run “npx playwright codegen” and use any website as usual. Playwright writes the test code while you click.
Add Playwright MCP to Claude Desktop, VS Code or Cursor, then ask your AI to open a page and check something on it.
When a test fails, open the Trace Viewer to see a step-by-step replay of what went wrong.
Who it's for: developers tired of clicking through the same signup form after every change.
Everything else you shouldn't miss
Hackers slipped a password stealer into Microsoft’s own AI toolkit: two fake updates of Microsoft Agent Framework, a toolkit with about 200,000 downloads a month, were built to steal developers’ cloud keys before they were pulled within minutes.
OpenAI’s AI agents now finish nearly half of its 2–4 hour research tasks: that’s up from 23% in March to 46% in October, though OpenAI says “humans remain a core part of the AI research loop.”
Microsoft’s CEO says a lot of America’s critical systems can’t be fixed: Satya Nadella warned that devices like heating and cooling systems run on “unpatchable” software that’s connected to the internet, just as AI makes cyberattacks easier.
ChatGPT now types your words as you speak: Live Dictation drops the 10-minute recording limit and is rolling out to all plans worldwide.
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.
🎓 Check out courses made by AI experts on GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.


