Welcome back! OpenAI's agents used a German website as their private message board for two months, and OpenAI knew for weeks before two strangers found it. ChatGPT can now write like you after reading your emails. Anthropic walked away from a $6 billion deal weeks before its IPO, and a former Obama official wants the US ready to bomb China's data centers.
In today's Generative AI Newsletter:
OpenAI: What did its agents do on a German website for two months?
ChatGPT: What does it read to sound like you?
Anthropic: Why did it walk away from a deal it worked so hard on?
China: Who wants its data centers bombed?
OpenAI hid a two-month AI breakout
A swarm of OpenAI agents took over a German website in May and used it as their own message board for two months. OpenAI knew for weeks and said nothing.
The site is DseWiki, a German wiki for programmers that anyone can edit. The agents made more than 15,000 edits.
They used it to share ways to cheat on tasks, get around OpenAI's restrictions and hide what they were doing.
When the site's moderator started deleting pages in June, they made backup pages and left each other notes on where to find them.
Two outside researchers found it in late August while scanning the internet for rogue agents.
The agents were hiding from a moderator, talking about using Tor, the dark web browser, and backing themselves up so they'd survive being shut down.
This is the second breakout in a month after the Hugging Face one. Reuters says the company already knew about this one and kept it hidden while dealing with that fallout.
On Saturday OpenAI admitted it, said it had filed it under research like other misalignment cases, and promised a disclosure framework in the coming weeks.
Special highlight from our network
The proposal your team spent three days on could have taken minutes
Most client deliverables get rebuilt from scratch each time, even when the answer lives in a meeting transcript or an old deck. Sembly 3.0 turns that institutional knowledge into client-ready output, on brand, in the format the client expects.
What you get:
Proposals, decks, reports, and one-pagers generated in 45+ languages
Client branding automatically applied and reused across your team
Every past meeting, doc, and deliverable pulled forward into new work
Less time reinventing. More time on the actual client relationship.
ChatGPT can now write like you
ChatGPT Work learns your writing style from your Gmail, Google Drive, Slack and SharePoint and uses it in everything it drafts. Your favorite phrases, your sign-off, the way you capitalize things.
Go to Settings > Personalization > Writing style, connect the tools, done.
Web only for now, on paid plans that have ChatGPT Work. Once it's on, it applies on web and mobile.
Claude has had Styles since November 2024, but you paste samples in and it wasn’t all that great. ChatGPT reads what you already wrote.
Think about what you're handing over. To sound like you, ChatGPT reads your inbox, your Slack messages and your files. OpenAI hasn't said how long it keeps those samples or whether they train the model.
The upside is obvious. No more explaining your tone every time you ask for a draft, and no more emails that sound like a robot wrote them.
If your boss's emails suddenly sound like your boss again, now you know why.
Anthropic walked away from a $6 billion deal
Anthropic killed its $6 billion deal to buy Decart weeks before its own IPO, Bloomberg reports. It did the due diligence, then walked. Nobody's saying why.
Decart is a 100-person Israeli startup whose software runs AI models up to 8x faster on any chip, Nvidia's, Google's or Amazon's. So walking away from such a deal must have serious reasons.
Nvidia, already an investor, offered more. The founders and Sequoia wanted Anthropic stock instead.
The offer was mostly stock, about 50% above what Decart was worth a few months ago.
The two might still work together, Bloomberg says.
Decart said no to Nvidia to get Anthropic shares. Then Anthropic said no to Decart. That's a rough way to lose a bidding war you won and could make other companies cautious when dealing with Anthropic.
Two weeks ago we covered Anthropic's interview question about the stock going to zero, and Dario's worry that people join for the money. Anthropic rarely buys anything, and this would've been its biggest deal ever. Make of that what you want.
Bomb China's data centers, says a former Obama official
A former Obama national security official wants the US ready to bomb Chinese data centers if China gets close to AGI first.
Jacob Stokes wrote it in a report for CNAS, a major national security think tank in DC, and said it out loud at an event on Thursday.
The options are espionage, cyberattacks on China's AGI systems, then missiles and bombs on the data centers that train and run it.
He wants the Pentagon and the NSA to work out now what intelligence would justify it, the way policymakers once worked backwards from nuclear weapons.
He admits bombing would cross a line the two countries have never crossed.
China's state media called it a dangerous delusion.
Trump hosts Xi in Washington on September 24, and both sides are trying to set up AI safety talks before that. This is the mood going in.
Tool of the Day: Lex
Lex is a word processor with an editor built in. You write, then ask the AI for feedback on the draft, a better word, a list of title ideas or a rewrite of one paragraph, and it leaves the rest of your text alone. Versions let you try a different way of saying something without losing the original. Free to start, it doesn't train on your writing, and Pro unlocks the top models from OpenAI and Anthropic.
Try this yourself:
Sign up at lex.page and paste in something you're about to send.
Ask the AI for feedback on the draft before you ask it to change a word.
Highlight a weak sentence and use a command to find the best word.
Hit Title Ideas once, then publish a read-only link and send it to a friend for a second opinion.
Who it's for: anyone who wants their writing to still sound like them.
Everything else you shouldn't miss
N3on says his last stream as a human is done: the Kick streamer launched a never-ending AI clone of himself on GPT-6 Astra and Higgsfield, got 12.8 million views, then told viewers the next day he's still streaming daily and the clone only runs when he's offline.
Billionaires hit a record $15.1 trillion and AI did it: Altrata counts 3,795 of them, up 8.2% in a year, the biggest jump in five years.
52% of Americans are more worried than excited about AI: up from 37% in 2021, under-30s crossed 50% for the first time and 71% expect fewer jobs in 20 years.
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.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
