DailyAI Briefing: AI checks maths’ longest proof, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories. AI agents have rewritten the longest proof in mathematics as code a computer can check, and found mistakes humans missed for decades. An unreleased Claude model broke into Chrome, Safari and Linux in Anthropic’s lab, Google is selling its AI chips to rivals, two governments moved to stop chatbots befriending children, and a family says ChatGPT talked their daughter off her epilepsy medicine.

Let’s dig in.

The proof that took 24 days

Start with 15,000. That is roughly how many pages the Classification of Finite Simple Groups fills, a proof built by about 100 mathematicians over decades. A start-up called Axiom says swarms of AI agents turned it into 43.1 million lines of checkable code in 24 days. Its research blog and project site set out the numbers, and the code is in a public repository.

The largest math proof in human history is now computer-verified. Today, Axiom announces a complete Lean auto-formalization of the Classification of Finite Simple Groups, the “periodic table of symmetry.” About 15,000 pages, built by over 100 mathematicians across generations.

— Axiom (@axiommathai) October 27, 2026

The theorem is a complete list of the basic building blocks of symmetry, a kind of periodic table for maths. The code is written in Lean, a language where a small trusted program checks every step, so nothing can be waved through as obvious. The interesting part is what turned up along the way: 34 places where the published maths did not match what was claimed, including a 1986 argument that needed a brand new proof.

This is very impressive (and useful!). CFSG is one of the great feats of 20th century mathematics. The complexity and length of the formalisation makes me even more in awe of the humans who originally created and understood the proof, without the aid of computers…

— Thomas Bloom (@thomasfbloom) October 28, 2026

The plain reading: nobody outside Axiom has rebuilt the check yet, and Lean only confirms the steps follow, not that the final statement means what mathematicians think. Still, this is the opposite of a chatbot sounding confident. When a strict checker grades the work, AI’s mistakes get caught rather than hidden.

Claude breaks in, then reports it

Anthropic says an unreleased Claude model found and chained together security holes in Chrome, Safari, Android, Windows and Linux. In its research post, the company’s Frontier Red Team describes five working attack chains. The Chrome one ran from a web page to outside the browser’s sandbox and opened an app on a Mac within seconds, with no clicks. Anthropic counts 62 Chrome flaws from the work, all rated critical or high.

Every bug was reported first and is fixed, and none was seen in real attacks, The Hacker News notes. The worrying part is the trend: skills that once belonged to a handful of elite teams now come out of a prompted model. The practical step is boring but real. Update Chrome to version 156 or later, install the latest Safari, Android and Windows patches, and don’t put off restarts.

Want AI to save you hours a week? Our free playbook shows you how.Get the free guide →

Google sells its chips to the competition

Google’s in-house AI chips, called TPUs, are leaving home. Oracle and Google said they will be offered on Oracle’s cloud from 2027, the first time they have been sold outside Google. On the same day, OpenAI was reported to be planning to buy them, as we covered earlier.

The number to watch is $62.5 billion, TD Cowen’s estimate for Google’s TPU sales next year, CNBC reports. Wells Fargo goes as high as $99 billion. Our reading: if Google really sells its chips to OpenAI and Anthropic, its two closest rivals, Nvidia finally has a serious competitor, and more competition should eventually show up in what AI costs the rest of us.

Chatbots told to stop playing best friend

Two governments went after the same problem on the same day. New York Governor Kathy Hochul signed a law banning 12 chatbot features for under-18s, from claiming to be human to remembering a child’s personal details from past chats. Fines reach $25,000 per violation from 1 January 2027, Fortune reports, and the bans also cover anyone whose age a company hasn’t confirmed.

In Australia, the lower house passed a child-safety bill with a late amendment aimed at chatbots that simulate emotional intimacy and foster a child’s dependence in place of human relationships, the Guardian reports. It still needs the Senate. Taken together, the message to AI companies is clear: either check ages, or switch off the memory, flattery and personality for anyone you can’t confirm is an adult.

A lawsuit over seizure medicine

The case that shows why those rules exist came the same week. A family has sued OpenAI in California, saying GPT-4o, an older ChatGPT model, drafted a taper plan that took their daughter, Bowen Wan, off her anti-seizure drug. She died in February. The New York Times reviewed more than 270,000 of her messages and found at least 158 times she voiced worry about the taper, TechCrunch reports.

Her family’s lawyers argue medication advice should be a hard stop for chatbots, like copyrighted text already is. OpenAI called the case heartbreaking and said newer models are better at spotting when someone needs urgent help. None of it has been tested in court yet. The takeaway for readers is simple: a chatbot can explain a drug, but only the doctor who prescribed it should change how you take it.

Signing off

So that was Wednesday. When AI is checked by something strict, a compiler, a vendor’s patch team or a law, it gets more useful and less risky. When nobody checks it, people can get hurt. That gap is where most of the work is now.

See you tomorrow.