Welcome back! OpenAI built a version of GPT-6 for lawyers and the legal AI startups it competes with are already building on top of it. Anthropic revamped projects in the best update they had in a while. Three researchers used Claude Opus 5 to break into OpenAI's private code and got paid for it. Trump's state dinner for Xi Jinping has three AI bosses on the guest list.
In today's Generative AI Newsletter:
OpenAI: What happens to legal AI startups now?
Anthropic: What changed in your workflow?
Hacktron: How much does breaking into OpenAI get you?
The White House: Which three AI bosses made the guest list?
OpenAI came for legal AI and its rivals signed up
OpenAI released Astra for Law, a version of GPT-6 Astra tuned for legal work. It isn't a new model. It's the same one wrapped in a legal search index, legal writing instructions and controls for confidential client files.
The index covers US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs.
What OpenAI is claiming:
On 200 questions from Vals AI's Legal Research Bench, it passed the correctness check 54% of the time against 38.7% for Astra with plain web search.
It found 24% more relevant cases and up to 54% more relevant passages from the right opinions.
Harvey and Legora, the legal AI startups OpenAI is supposedly competing with, are building on top of it as API customers.
Anthropic launched Claude for Legal in May and Google shipped Gemini for Legal in August. OpenAI hired Ironclad's founder in June to run this one.
Special highlight from our network
Learn what’s being built next
Prepathon 2026 is two days of real conversations with folks from Google, Amazon, OpenClaw, WooCommerce, Elementor and 20+ more.
AI, development, ecommerce, agencies, infrastructure, all of it.
Day 1 has a dedicated session with OpenClaw's DevRel manager, which we're pretty excited about.
No fluff, no keynote theater. Just people explaining what they're building and what's coming next.
It's free. Come sit in and get practical insights.
Once in a lifetime opportunity
Own a piece of the AI go-to-market accelerator
300+ AI brands already pay GenAI Works to reach 14M+ people who follow AI on purpose. Nvidia, Oracle, Google, IBM and Notion are clients.
Revenue grew 2.5x last year. $3.5M already in the funnel.
Invest from $1,000, up to 20% bonus shares.
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.
Anthropic put one Claude in charge of the others
Anthropic rebuilt Claude Projects around a coordinator that hands work to other Claudes.
You describe what needs doing. Claude splits it into jobs and opens a thread for each one, then reviews what comes back and gives you a single finished result.
What that looks like underneath:
Each thread is a full Claude Code cloud session with its own branch and its own copy of your repo.
Threads keep running after you shut your laptop, and you can steer them from your phone.
Every thread writes into one shared memory, so a release date that changed or a service nobody touches without approval carries across all of them.
Two threads editing the same file is still a merge conflict, same as any other pull request.
Every thread counts as a full session, so a project running several at once burns your limits much faster. Anthropic got hit with a class action nine days ago over how it counts Max usage.
What changes for you is the unit of work. You stop opening a session per task and start handing over a goal, then checking the overview for the threads that are blocked or ready to review.
It's in beta for some Pro and Max subscribers who already use cloud sessions and don't have projects on web or desktop. Everyone else gets a waitlist.
Someone used Claude to break into OpenAI
Three researchers at Hacktron AI used Claude Opus 5 to get inside OpenAI's private code in July, and they've only now published how.
They uploaded a rigged image to OpenAI's public help forum
The software that handles uploads ran the code buried in the file instead of showing a picture
That gave them the run of the forum's server
A second flaw, this one in OpenAI's own login, let them into the ChatGPT and Codex accounts of anyone who had ever signed into that forum. Employees included. They reported it to OpenAI's bug bounty within hours.
Then they proved the reach.
One employee's Codex was wired into OpenAI's GitHub, so they told it to open a harmless pull request in the internal repo and stopped testing.
Claude Opus 4.8 couldn't make the exploit work reliably across several sessions. Anthropic shipped Opus 5 that evening, they handed it the same problem, and it worked within hours. Start to finish took under 72 hours.
The wider campaign, which also hit Slack, Meta, Zoom and Shopify, cost under $3,000 in tokens. OpenAI fixed its side the same day and paid them $6,500.
AI CEOs are on equal footing with world leaders
Sam Altman, Jensen Huang and Tim Cook are on the guest list for Trump's state dinner for Xi Jinping on September 24. OpenAI's head of global affairs confirmed Altman is going.
Cook left the Apple CEO job on September 1 and took the executive chairman title, which at Apple means government liaison, so new CEO John Ternus stays out of political rooms.
All three get seats at a head-of-state dinner in the same month their own industry asked Washington to slow AI down and both the American and Chinese presidents called the worry a hoax.
Bernie Sanders, one of the most influential American politicians, says he hopes the two presidents use the meeting to negotiate a treaty pausing advanced AI development.
Top 5 tools of the week
AI resume builder, ATS resume checker & career tools. Trusted by 9M+ job seekers. Build a resume & cover letter, and land your dream job fast
Krea AI is a real-time creative platform for generating and enhancing images, videos, and 3D visuals.
Grammarly is the world's most trusted AI writing assistant, helping millions polish grammar, tone, and clarity across email, docs, and apps.
Tavus creates hyper-personalized AI videos featuring your own digital replica, complete with real-time conversational abilities.
Replit is a browser-based coding platform with Agent, an AI that builds, debugs, and deploys full applications through natural conversation.
Everything else you shouldn't miss
OpenAI is closing in on a second $1 million math problem: it won't say which one, mathematicians think it's the Hodge conjecture, and more than 1,000 of them signed a letter calling the approach research misconduct.
Bernie Sanders wants 20-year sentences for building superintelligence: he shared a stage with Steve Bannon to argue AI is more dangerous than nuclear weapons.
ChatGPT ads can now talk back: OpenAI is testing Sponsored Agents with US advertisers, so the ad becomes a conversation you can ask questions in.
The best coding agent still fails most of the time: on Real-SWE, which tests agents against private codebases, top scorer Fable 5.1 failed more than 60% of its attempts.
Firefox's AI mode now runs on Mistral: Smart Window added Mistral Small 4 and launched in France, with zero data retention on Mistral's side.
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.
