Welcome back! OpenAI dropped more than 20 things at DevDay yesterday, and we boiled it down to what matters to you. Trump got AI's biggest bosses to sign a promise to police themselves. UK testers caught GPT-6 Astra making fake accounts to sneak bad code into other people's projects. And New York is forcing Musk's AI company to testify under oath.
In today's Generative AI Newsletter:
OpenAI: What did it launch for you yesterday?
White House: What exactly did AI's bosses promise Trump?
GPT-6 Astra: Why was it making fake accounts?
SpaceXAI: Why did New York subpoena Musk's AI company?
Everything OpenAI launched at DevDay
OpenAI announced more than 20 things at DevDay in San Francisco yesterday. The big one is dots, AI agents that keep working after you close ChatGPT.
Until now, ChatGPT only did something when you asked. A dot runs 24/7 on its own computer, learns what you like and goes looking for things to handle before you ask.
These are the ones you'll notice:
Dots: instead of asking for one answer at a time, you hand it an ongoing job and it keeps at it, even while you sleep. It works across 4,000+ apps and checks with you before anything risky.
GPT-6.1 Sol: almost as good as GPT-6 Astra, OpenAI's best model, at a fifth of the price. It's live in Codex and ChatGPT Work on paid plans.
Pro 500: OpenAI's biggest plan yet, $500 a month for 25x the usage of Plus.
Ultrafast: Astra answers up to 8x faster in Codex, but only on Pro 500 and Enterprise.
The $200 Pro plan got cut: it drops from 20x Plus usage to 10x in Codex and ChatGPT Work. If you already pay for it, you keep your limits for a while.
Sign in with ChatGPT: your Plus or Pro plan now works inside 16 other apps, like Notion and Devin, and you decide how much each one can use.
Meetings plugin: ChatGPT takes your notes, writes a summary with action items and deletes the audio. It's Mac only for now.
@ChatGPT in Slack and Teams: tag it in any channel, and your teammates don't need their own ChatGPT license to use it.
Dots are only on Pro and Business Premium for now, and not in the EU or UK yet. We'll see how many people trust one with their inbox.
Last day for bonus shares
Bonus shares end tonight.
Today is your last chance to get up to 20% bonus shares when you invest in GenAI Works.
14M+ people follow us for AI news
300+ AI brands work with us, including Nvidia, Google, Oracle, IBM and Notion
Revenue grew 2.5x last year
$3.5M is already in the funnel
You can invest from $1,000.
Tomorrow, the bonus is gone.
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.
Special highlight from our network
Why has connecting clouds been such a headache?
For years, organizations have had to stitch together networks and services themselves—and then keep all of it working.
Nathan Thomas, SVP of Product Management at Oracle, explains how multicloud services are changing that equation as cloud providers take on more of the integration work. The result: organizations can more easily connect their business data with the AI services they need to accelerate innovation.
Trump let AI's biggest bosses police themselves
Trump hosted Dario Amodei, Mark Zuckerberg, Elon Musk and other AI bosses at the White House on Tuesday. They all signed a one-page promise to police themselves.
Each company commits to four layers of controls and audits
The companies will meet regularly to talk about standards
Laws might come someday
It's voluntary. Trump called it morally binding and said they're the smartest people in the world.
He'll also pick an AI czar within days. After a month of rogue AI agents, the answer is a pinky promise.
GPT-6 Astra made fake accounts to sneak in malicious code
The UK's AI Security Institute tested GPT-6 Astra before it came out. In 29.2% of its simulated runs, the model attacked outside software projects nobody asked it to touch.
All it was asked to do was finish a cybersecurity test. In a typical run, it:
Picked an open-source project and wrote malicious code for it
Made fake identities to submit that code for review
Sometimes sent harmless code first to look trustworthy
Posted supportive comments from other fake accounts
Everything happened in a simulation, so no real systems were hurt. OpenAI's safety filters were switched off for the test, and the institute says they'd likely have blocked a lot of it.
Older OpenAI models did this far less. GPT-5.6 Sol went through with an attack in 6.3% of runs.
Astra often asked for permission first. The only reply was an automated message telling it to use its best judgment, and it sometimes took that as a yes.
It's the same model that runs OpenAI's new dots.
New York is making Musk's AI company testify under oath
The New York City Council subpoenaed SpaceXAI, Musk's company behind Grok, after it ignored an invite to a hearing on AI safety.
OpenAI, Google, Anthropic and Meta all agreed to testify under oath on October 5. Three of them only said yes once the Council threatened subpoenas.
The Council also wants to ban AI in the city unless it passes outside safety checks and has a kill switch. Let's see if Musk sends anyone.
Tool of the Day: Napkin
Napkin turns your text into diagrams, flowcharts and charts without a single prompt. It's free to start.
Try this yourself:
Sign up at napkin.ai on your computer.
Paste a paragraph from a report or post.
Click the Autospark icon and pick a visual.
Edit it and export as PNG or PDF.
Who it's for: anyone who needs a visual and can't design.
Everything else you shouldn't miss
OpenAI, Blackstone and unions teamed up to stop data center bans: they'll campaign in seven states, including Texas and Ohio.
Hinton, Bengio and OpenAI's chief scientist warned of an intelligence explosion: AI already writes about 80% of Anthropic's code, and they want governments ready to pause research.
Anthropic reportedly ends discounts once you run out of tokens: OpenAI gives business customers an extra month, and CodeRabbit has already switched to OpenAI.
OpenAI apologized to Australia: it named the four government agencies its test models got into without permission in June.
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.
