← Home AI in 15

AI in 15 — November 11, 2026

November 11, 2026 · 15m 57s
Kate

Six million lines of mathematical proof, written by AI agents in four weeks, covering the encryption code that protects billions of private messages. Google DeepMind says that job used to take expert teams years.

Kate

Welcome to AI in 15 for Wednesday, November 11, 2026. I'm Kate, your host.

Marcus

And I'm Marcus, your co-host. Happy Wednesday, Kate. Today the machines are checking our homework, and for once I'm in favor.

Kate

You've clearly never seen my homework. Here's what's coming up. DeepMind's AlphaProof Code formally verifies the code behind the Signal protocol.

Kate

Google releases Gemini 4 Carbon, its new top model, at half price for three weeks.

Kate

Anthropic's IPO finally has a price range.

Kate

The Fireworks AI breach reaches Notion, and a zero-retention promise falls apart.

Kate

Plus a reality check on AI bug hunting, Apple's plan to turn Macs into AI servers, and a video game built from one prompt.

Kate

Marcus, let's start with the big one. Formal verification. What is it, and why should anyone outside a math department care?

Marcus

Normal testing tries a lot of inputs and hopes it catches the bugs. Formal verification is a mathematical proof that the software does what its specification says, for every possible input. It's the gold standard, and it has always been brutally expensive. You see it in aerospace and a handful of niche projects, and almost nowhere else.

Kate

And DeepMind pointed AI at that problem.

Marcus

AlphaProof Code is a framework where several AI models work together. The division of labor is the interesting part. Humans decide what needs proving, and the agents do the proving. The first target was the compiled code behind the Signal protocol, the encryption used in Signal and Google Messages. DeepMind stresses that the proofs cover the code that actually runs on devices, not just the design on paper.

Kate

What kind of thing did it prove?

Marcus

One example: your messages stay protected before and after a device is temporarily compromised. To get there, agents running on Gemini Carbon wrote more than six million lines of Lean, which is a language for proofs a computer can check. About one percent of the codebase was flagged for human review.

Kate

How do we know the proofs mean anything?

Marcus

They ran a nice test. Researchers took two bugs that had already been fixed and put them back into an isolated copy of the code. Both times, the system produced a machine-checked proof that the buggy version was insecure. And Signal's CTO, Ehren Kret, called translating live code into valid Lean theorems "a major breakthrough."

Kate

I can hear a "but" warming up.

Marcus

A couple. This was the top story on Hacker News, and the skeptics made good points. One doubted that a library with a history of bugs could be verified and turn up no new ones. Others noted the method checks the code against the spec, and nobody is proving the spec itself is right. If the humans wrote down the wrong property, you get a flawless proof of the wrong thing. And AI-written Lean can cheat if nobody is watching.

Kate

So who checks the checkers?

Marcus

That's the question. The good news is DeepMind open-sourced the proof artifacts and published a paper, so outsiders can actually go and look. I'd want an independent team to reproduce it before calling it done. But if the cost really falls from years to weeks, verified software stops being a luxury. One commenter predicted that in a few years, running unverified code will feel as reckless as using Telnet.

Kate

For anyone under forty, Telnet is how we used to log into computers with no encryption at all.

Marcus

And we thought it was fine.

Kate

Quick hits. The model behind that work is now out. Google released Gemini 4 Carbon. Marcus, where does it sit?

Marcus

At the top of Google's lineup, above Gemini 4 Argon, which Google still recommends for everyday use. Carbon spends more tokens exploring a problem and checking its own work. It's aimed at long multi-step agent tasks, cybersecurity and scientific reasoning. Koray Kavukcuoglu says it's "built for tasks that test the limits of a model's intelligence."

Kate

And the scorecard?

Marcus

These are all Google's own claims. First place on CursorBench, Perplexity's WANDR benchmark and Zapier's AutomationBench, at less than half the cost per task of the next-best outside model. In internal security tests, a hundred percent recall on scanning and ninety-five percent precision on triage. Ninety-two percent on an honesty benchmark, where Argon got seventy. And Google admits that resistance to prompt injection still needs work, which I appreciate them saying out loud.

Kate

Now the price, because there's a catch.

Marcus

Two dollars per million input tokens and ten per million output, the same as Argon. That doubles after November thirtieth. It's a three-week introductory offer, and it's clearly aimed at developers who have settled in with Claude.

Kate

A free sample at the supermarket.

Marcus

With a very large supermarket behind it. Early reaction is mixed. Gemini Pro subscribers are annoyed that in the app it's only for the pricier Ultra tier. One benchmark runner said it matches Anthropic's Fable 5.2 but is slower and more costly, and another said it beat Fable on critiquing a paper draft. Simon Willison ran his pelican-on-a-bicycle test, and a clip from his testing even made Google's launch video. Google's Antigravity coding tool also gets a preview swarm command, which coordinates teams of agents from one prompt.

Kate

Anthropic's IPO. Yesterday we told you the company sent bankers numbers and then retracted them. Now there's a real filing.

Marcus

Filed Tuesday, Nasdaq, ticker A-N-T. According to secondary reports, since the Forbes and Wall Street Journal pages were out of our reach, it's eight hundred and twenty-five million shares at seventy to seventy-eight dollars. That's up to about sixty-four billion from the offering, or roughly ninety billion including the private placement. That would be the largest IPO ever.

Kate

So slightly fewer shares and a slightly lower top than the retracted version. And the valuation?

Marcus

Headlines say about one point three trillion dollars, but it depends on how you count shares. Using shares outstanding at the top of the range, it's closer to one trillion. Fully diluted, estimates run from about one point two six to one point four. Trading reportedly starts around November nineteenth or twentieth.

Kate

Is that number up or down from expectations?

Marcus

Both. It's above the nine hundred and sixty-five billion from the confidential filing in June, and below the one and a half to two trillion people were floating in late October, before the AI stock sell-off. The Financial Times reports bankers are weighing a price above the range because early demand is strong. One outlet says the founders will each sell around two hundred and fifty million dollars of stock. We couldn't confirm that, so file it under maybe.

Kate

And the debate?

Marcus

One side says it's justified if Claude stays even slightly ahead. The other asks where the winner-takes-all outcome is when strong open models exist. Both are reasonable. The useful part is that public investors finally get to put a price on a frontier lab, and OpenAI will be measured against it.

Kate

Security. Yesterday we mentioned the Fireworks AI breach. It's grown, and Notion users are caught in it.

Marcus

Quick recap. Fireworks hosts and runs open models for other companies. It's now had two intrusions. On October sixth, someone used stolen internal credentials. Then in late October, a breach at a third-party cloud provider gave attackers a way into Fireworks' development machines. The November ninth update confirms that for a limited set of customers, attackers reached account and billing data, secrets, datasets and incomplete fine-tuned model weights. So the Forbes detail we flagged yesterday as unconfirmed is now in the company's own bulletin.

Kate

And Notion?

Marcus

Notion says fewer than one percent of its workspaces were affected. Chat transcripts from Notion AI, used for a feature called Insights, were exposed, though without labels identifying the user or workspace. The awkward part: Fireworks had contractually committed to zero data retention. Notion says it later learned data was being kept from the endpoints it used.

Kate

Wait. So the contract said "we don't keep your data," and they kept the data?

Marcus

That's Notion's account. Fireworks says its zero-retention inference platform wasn't affected, so the two versions don't line up neatly yet. Palo Alto's Unit 42 is investigating. Notion cut ties on November third and switched Insights off.

Kate

What should people do?

Marcus

If you ever pasted a password or API key into Notion AI, rotate it. And for companies, the lesson is that zero retention is a promise, and promises need auditing. The weak point was a smaller vendor two steps down the chain, and customers had no way to see it.

Kate

More security, and this one is a reality check. An independent firm went through the numbers on Anthropic's Project Glasswing.

Marcus

Glasswing is Anthropic's program using its Mythos models to find vulnerabilities. VulnCheck counted four hundred and forty-two CVEs, the public vulnerability IDs, attributable to Anthropic in seven months. That's about half a percent of this year's total, and only two are known to be exploited in the wild. Anthropic's own ledger lists nearly thirty-eight thousand findings, with two hundred and seventeen marked fixed.

Kate

Thirty-eight thousand found, two hundred fixed. That's my inbox.

Marcus

And it's getting worse. Since September, findings grew by about eleven thousand eight hundred, and fixes grew by fifteen. Then there's curl, the widely used data transfer tool. Of thirty-four Anthropic reports published so far, nineteen weren't vulnerabilities. That's a false positive rate of about fifty-six percent. Claude also rated over ninety percent of findings critical or high, where maintainers said about half.

Kate

Is there a defense?

Marcus

A real one. The developer of Fil-C said all nine reports he received were genuine bugs of exceptional quality. An engineer at a big tech company said they've been patching "company-ending" bugs daily since Mythos launched. And critics say public CVEs are the wrong yardstick, because most finds happen inside partners' private code. That last claim is hard to verify from outside, which cuts both ways. Finding bugs is now cheap, and the triage and fixing lands on maintainers who are often volunteers.

Kate

Apple. A leaker spotted an app called Harbor that turns your Mac into a private AI server.

Marcus

Aaron Perris found it, via 9to5Mac. It's in private preview. You browse, download and run open models without touching the command line. The headline feature is that you can connect two or more Macs over Thunderbolt 5, and Harbor pools their memory so you can run models too big for one machine.

Kate

So my old Mac mini finally gets a purpose.

Marcus

If it has Thunderbolt 5, yes. Oddly, it's coming out under Claris, the Apple brand best known for FileMaker. There's no release date. It fits Apple well, though. Macs have lots of unified memory, and Harbor puts that to work in a space that hobbyist tools like Ollama and LM Studio serve today. Keeping your data on hardware you own looks even better after the Fireworks story.

Kate

For the developers. Google's Chrome team released a package teaching coding agents to write modern web code.

Marcus

It's called Modern Web Guidance, it's open source, and it covers about two hundred and fifty web platform features. It plugs into Claude, Codex, Cursor, Gemini, Grok and Kimi. The idea is that agents learned from years of old code, so they reach for heavy JavaScript where the browser can now do the job natively. There's also a command that scans your codebase for outdated patterns.

Kate

Sounds helpful. What's the objection?

Marcus

Edge, Firefox and MDN are listed as stewards, but a top Hacker News comment listed recommended features that work in Chrome and not in Firefox or Safari. People still remember "best viewed in Chrome." Whoever writes the style guide for the agents has a lot of say over how websites get built, so I'd like to see the other browsers actually editing it.

Kate

And something viral. A first-person shooter called Saltline, built by Claude Fable 5.2, the model we covered yesterday.

Marcus

It runs in the browser as a Claude artifact. Per Simon Willison, citing a post from Anthropic's Felix Rieseberg that we couldn't verify directly, the prompt was roughly "build me an FPS, photorealistic graphics, small in scope." The reviews: graphics impressive, gameplay mid, and miserable on a phone. Plenty of people also asked how many credits it burned.

Kate

"Graphics great, gameplay mid" describes half the games I've bought.

Marcus

At full price, too. One more for the road. A study from ActivTrak found that AI users' workdays run up to ninety minutes longer than non-users'. Someone called AI an e-bike for the mind. Apparently we're using the e-bike to commute further.

Kate

One to watch: Anthropic's final IPO price. If bankers go above that seventy to seventy-eight dollar range before next week's debut, it tells us how much appetite survived the sell-off.

Marcus

Agreed. I'd also keep an eye on outside teams trying to reproduce those Signal proofs.

Kate

That's your AI in 15 for today. See you tomorrow.