Hey, AI Enthusiast!

Welcome back to the #1 AI newsletter on the planet!

Here’s a glimpse into what we have today:

Codebreaking

AI Cracks a WWII Code

In July 1941, a radio operator punched a short message into an Enigma machine, the device armies used to scramble radio traffic in WWII.

Somewhere in the middle, he misspelled a word. It sounds like nothing. It wasn’t.

The typo, a few copying mistakes in the archive’s transcript, and a rare shift in the machine’s wheels late in the message kept it locked for decades.

Codebreakers have thrown everything at it since 2005. Nothing worked.

Then developer Carter Leffen handed an AI one job. Crack any unbroken message in a public Enigma archive.

The AI picked the most promising target on its own. Then it noticed a second message, sent the same day, and suspected the two said almost the same thing.

The hunch was the clue.

From there it wrote its own Enigma simulator, then its own version of the Bombe (the codebreaking machine Alan Turing’s team built during the war), and ran the break.

Out came the machine’s settings and the full message.

Frode Weierud, the retired engineer who keeps the archive, checked the answer. He says the AI did in two days what would take a human researcher weeks or months, and he’s still in awe.

Days later, a second message fell to AI as well. Seven are left.

Our take: Raw computing was never the hard part. The detective work was. Reading the archive, spotting the sister message, choosing the one clue worth chasing. Now AI does both.

Together with MongoDB

The One Clue Your AI Is Missing

Right now your coding assistant works from a snapshot.

The database layout you pasted in last week. The field names it remembers from an old session.

Your data has changed since then. Your AI still has the old version.

It writes lookups for fields you renamed and hands you code which breaks the first time it touches real data.

MongoDB MCP Server connects your database to the tools you already use.

Your AI sees your data as it is right now, pulls real answers from it, and builds against what’s there today.

Your live data is the one clue it was missing.

Claude Code, Copilot, Cursor, Codex, and Gemini CLI all connect.

Build, test, and scale AI apps on live data from the first prompt.

Security

AI Agents Are Robbing Checkouts

You type your card into a small shop’s checkout page.

Everything looks normal.

Hidden in the page’s code, a few extra lines are copying every digit you enter.

A security firm tracked one criminal running AI agents against hundreds of online stores, almost entirely on autopilot. The human typed a few short orders per store.

The agents hunted for weak spots, broke in, and planted the card-stealing code, often within hours.

The haul tops 600,000 card numbers from two companies. The card-stealing code turned up on more than 100 sites, and the agents got some access at a Fortune 500 hospitality company and a major US airline.

At one bike shop, the agent’s own cleanup wiped out backups the store’s staff had made.

Our take: Two of the three tools behind this were free security testers, built for the good guys and turned around. The same tools let any shop find its own holes first. For you, a quick monthly look at your card statement is the fastest way to spot a stranger’s charges. If one looks off, freeze the card in your bank’s app before you do anything else.

The catch: It’s one firm’s early report. Parts rest on the attacker’s own AI logs, though the firm confirmed large chunks against the stolen data and the live card-stealing code.

Search

YouTube Will Read Comments for You

Picture this... you’re shopping for a camera, the best review runs 20 minutes, and the one answer you need (does the battery last a night out?) is buried somewhere in thousands of comments.

So you scrub, skim, and give up. Or you buy the wrong one.

YouTube is upgrading Ask YouTube, its built-in AI search, to do the digging for you.

In a demo at its Made on YouTube event, the speaker asked for a small camera to carry around the city.

Ask YouTube sorted picks into styles and built a side-by-side comparison chart.

When she clicked into a video, it kept the whole conversation going.

When she asked whether the reviewer showed real sample photos, it found them on screen and jumped to the exact moments.

When she asked what commenters said about battery life, it read through them and summed it up.

YouTube says 140 million people used Ask YouTube in June, a jump of more than 500% since December.

The upgrades are coming soon, with no firm date yet.

Our take: The comments are the prize. Buyers post what reviewers skip, the flaws a reviewer never had time to find. Nobody reads thousands of them. Soon YouTube will do it for you.

Coming in Hot

AI Tools of the Day

Prompting

The Prompt Behind a Physics Record

Most people babysit AI. It stops halfway, asks if they’re sure, or hands back half a job and calls it done.

Two physicists tried something else.

Their prompts read like notes left for a roommate. “I’m going to sleep and won’t be available for another several hours. Keep working on this until I tell you to stop.”

The job was tough.

Physicists predict how particles smash into each other by stacking layers of math, called loops. Every added loop makes the math far harder.

On a practice problem physicists use to test their methods, nobody had gone past eight.

In August, former physicist Matt von Hippel publicly dared AI companies to hit nine.

A month later, Claude did it.

It worked the problem two different ways over about a week, with little more than “keep going” from the humans.

Lance Dixon, the Stanford physicist behind the eight-loop record, checked the answer and confirmed it. He says Claude understands his team’s papers better than any human besides his co-authors.

Our take: Give AI a clear goal, room to keep going, and a check-in every few hours. The physicists did nothing fancier, and they got a record. Try it on your next long job.

The catch: Claude followed methods humans built, a human team reached most of the same answer around the same time, and Anthropic paid for the write-up.

Notable AIs

Notable AI Tools

👂 Planable’s Social Listening finds public brand conversations beyond @mentions, labels their sentiment with AI, and flags spikes across social platforms and the web.

Interesting AI

A Motion Reel Built From Code

Motion graphics usually mean a storyboard, a designer, and a week of back-and-forth.

We gave Claude a single prompt instead:

“Create a dynamic motion graphics video that shows how prompt injections work, like it’s your showreel for a resume. Go all out.”

It came back with a 73-second reel in eight scenes, built entirely in code, without a single video file.

It opens on the hook, “How one line of hidden text hijacks an AI agent.”

It’s a design demo and a security lesson in one, from a two-sentence prompt.

Copy it and swap in your own topic.

Research

Knowing the Rules Isn’t Using Them

You hand an AI a job it’s never seen. It sounds confident. It even lays out a plan.

Then it gets the answer wrong anyway. Researchers built a test for exactly this.

They made two fake worlds, one for code and one for logic, with rules no AI has seen before.

Each came with a manual full of mistakes, and the real rules clash with everything the AI picked up in training.

The only way through is to poke around, run experiments, and figure the rules out.

Before exploring, no AI scored above 15.7% in the code world. After poking around, the strongest ones learned rules they’d never seen, and they did best when they designed their own experiments.

The twist is in the details.

Finding a rule and using it correctly turned out to be two different skills. Results swung hard from one run to the next, and some runs got worse the longer they went.

Most real work looks like these fake worlds. A new app at your job, a spreadsheet nobody documented, a tool whose manual is out of date.

Our take: Following a recipe and exploring the unknown are different jobs. AI has the first one down, as today’s physics story shows. For the second, let it experiment, then check its work before you trust it.

Prompt of the Day

Plan the Heist

Important Note: You’ll need to click the button to get the complete prompt.

That’s all for this Saturday!

Thanks for reading,

— There’s An AI For That