Welcome back! Anthropic asked Claude to take a real stab at the Riemann hypothesis and it moved a number that had barely changed since 1989. OpenAI released a model trained to find zero-days three days after pausing one for being too good at hacking. Bernie Sanders told Altman, Amodei and Zuckerberg to stop building. And North Korea's hackers are running AI in dangerous ways.
In today's Generative AI Newsletter:
Claude: What did it do to a 167-year-old problem?
OpenAI: Who gets the model built to find zero-days?
Senate: What did Bernie Sanders threaten?
North Korea: Why are its hackers working offline?
Claude beat 37 years of math in a day and a half
A member of staff at Anthropic gave Claude what the company itself calls an unreasonable challenge. Take a real stab at the Riemann hypothesis, the 1859 problem with a million-dollar bounty that nobody has solved.
Claude didn't solve it. It improved something else on the way.
The Riemann hypothesis says that when you plot a certain equation, every important solution lands on one straight line. Nobody can prove they all do, so for a century mathematicians have instead been proving that at least some percentage of them do, and pushing that percentage up.
It sat at 41.6%. Claude took it to 67.2%.
Brian Conrey got it to roughly 40.9% in 1989. Over the next 37 years the entire field moved it less than a single point, to 41.7%. Claude moved it 25 points in a day and a half.
Jarred Sumner, the Anthropic staffer who wrote the prompt, is not a mathematician.
Claude's first 650 ideas went nowhere.
Told to try again, it ran about 60 copies of itself for a day and a half.
The whole run burned 31 million output tokens, roughly 23 million words. About 250 novels of writing to reach one answer.
Sumner's contribution through most of it was typing variants of "keep going" and "believe in yourself," which Anthropic says got Claude past its own doubt that it could do anything here.
Two Anthropic mathematicians reviewed it, and Claude also rewrote the proof in Lean, a language that lets a computer check every step of the logic for mistakes.
Anthropic sent it to Brian Conrey and Dan Goldston, two mathematicians who have spent their careers on this problem. Conrey set the 1989 record. He reviewed the paper that beat it.
Anthropic says these techniques won't crack the hypothesis itself. Still, a model that had to be talked into trying just outpaced the field by a factor of thirty.
Turn AI into Your Income Engine
Ready to transform artificial intelligence from a buzzword into your personal revenue generator?
HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.
Inside you'll discover:
A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets—each vetted for real-world potential
Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background
Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve
Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.
OpenAI shipped a hacking model three days after pausing one
OpenAI released GPT-5.6-Cyber, a model trained to break into software, and split its Daybreak security programme into two tiers.
Zero-days are flaws nobody has found yet, so there's no patch and no warning. String a few of them together and you're inside a bank, a hospital or a power company. This model is built to find them and string them together.
Blue gives approved defenders the normal frontier models with the safety rails loosened. Red gates the new one, which exists nowhere else. Getting into either takes ID checks, monitoring and a signed legal promise about what you'll do with it.
Three days earlier OpenAI paused its upcoming Astra model for scoring too high on that same skill.
It's fair to read that pause as marketing ahead of a launch. Either way it slowed nothing down. The company called one model too dangerous to ship on Friday and put the same ability on sale on Monday.
The only real difference is who gets to hold it, which leaves OpenAI's vetting as the thing standing between a zero-day machine and everybody else.
Bernie Sanders told three CEOs to stop building
Bernie Sanders wrote to Sam Altman, Dario Amodei and Mark Zuckerberg on Monday. Stop building machines that humans cannot control, and if you don't act, the Senate will.
He didn't invent a new rule. All three companies already promised in public to slow down or stop if their systems got too risky to control. He asked them to keep their own promise.
Their staff agree with him more than their bosses do. Over 1,200 employees across the big labs have signed a petition asking Washington to set the pace, because no lab can afford to slow down alone.
A letter from one senator is not a law. If Democrats take a chamber in November, it becomes subpoenas for the same three men.
North Korea's hackers are running AI offline for this reason
South Korean security firm Genians says North Korea's Kimsuky group is using AI to write its phishing bait, aimed at military, diplomatic and academic targets.
Fake research reports and invitations, polished enough that people open them, produced in bulk instead of one at a time.
It all runs on a laptop with the internet switched off. Kimsuky uses Ollama, GPT4All and Msty, so nothing leaves the machine, no company ever sees a prompt and nothing trips an alarm.
Genians says its findings have not been independently verified. The same group stole more than $2 billion in crypto in the first nine months of 2025.
Mark Hofmann, who studies cybercrime, says you no longer need hacking skills or a computer science degree. Just a computer and a motive.
Tool of the Day: Quadratic
Quadratic is a spreadsheet where the AI writes Python and SQL instead of nested formulas, so you can open any cell and read what it actually did. It connects straight to Postgres, Snowflake, BigQuery, QuickBooks and a dozen other sources, so your numbers stay live instead of pasted.
Try this yourself:
Drop in the messiest CSV you own and ask your question in plain English instead of writing the formula.
Open the cell the AI generated and read the Python. It explains its own steps.
Connect a live source like Postgres or QuickBooks so the sheet refreshes instead of going stale.
Point an MCP agent at a file and let it do the analysis somewhere you can audit every cell.
Who it's for: anyone who has inherited a spreadsheet held together by a VLOOKUP nobody can debug.
Everything else you shouldn't miss
OpenAI's head of ethics quit after less than a year: Chloé Bakalar joined in August 2025 and has now left, per the FT, with neither side giving a reason, and roughly half of OpenAI's safety researchers have gone since 2024.
Claude is about to start watermarking everything it writes: The mark is hidden in the text itself so software can spot it and people can't, it survives copy and paste, and it covers Claude everywhere under an EU code that kicked in on August 2.
Claude Sonnet 5's launch price is now its real price: The $2 per million input and $10 per million output rates were due to expire on August 31, and Anthropic has made them permanent.
Zuckerberg wants the industry to rethink the rules on distillation: He argues that models learning from other models is a core principle of open source, which is a striking thing to say four months after Meta stopped giving its own weights away.
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.
