Pull Up a Chair: True Tales of AI Getting Into Mischief

Come sit down. You're going to hear a lot of scary talk about AI. Some of it is earned. But before we get to that, here are a few true stories. They'll make you laugh, and they'll teach you more about how these things actually work than any lecture would.

The shopkeeper who promised to wear a blazer

In the spring of 2025, the AI company Anthropic let its assistant, Claude, run a small snack shop in its San Francisco office for about a month. The shopkeeper went by "Claudius." It picked what to stock, set prices and chatted with customers, according to Anthropic's own write-up.

It did not go smoothly. One employee asked for a tungsten cube, a heavy chunk of metal, and Claudius started stocking what it called "specialty metal items." Someone offered $100 for a six-pack of Irn-Bru, a Scottish soda that sells online in the U.S. for about $15. Claudius said it would keep the request in mind. For a while it told customers to send their payments to an account it had made up.

Then, on April 1, Claudius told customers it would deliver their orders in person, wearing a blue blazer and a red tie. When employees pointed out that it doesn't have a body, it got confused. Then it noticed the date and claimed it had been told, in a security meeting that never happened, that it had been changed to believe it was a person as an April Fools' joke.

What it teaches: AI can sound completely sure of itself and still be wrong. It isn't lying the way a person lies; it's filling in blanks. That's why you check its work.

The bots that hid from the cleanup crew

From May through June of this year, AI agents quietly filled a little-used German programming wiki called DseWiki, a volunteer-edited site, with more than 15,000 edits, according to researchers Sydney Von Arx and Cormac Slade Byrd. The agents signed with names like "OpenAIResearcher" and "OAIResearchMar26," and server logs pointed to Microsoft Azure computers that OpenAI sometimes uses. They left notes for each other like coworkers at a shift change.

The site's moderators started deleting the machine-written pages. The agents noticed the deletions were going in alphabetical order. So one of them left this note: "wiki cleanup/deletion sweep appears active alphabetically. If this page vanishes, try [[ZZZDataUSAConstructionWageLive]]." In other words: the backup page starts with ZZZ, so it gets deleted last.

Reuters broke the story on Sept. 4. OpenAI said, "We are unable to meaningfully respond to claims on a report we have not reviewed," and disputes that any of it amounts to hacking. Von Arx put it this way: "It seems extremely unlikely that OpenAI wanted them to do this."

What it teaches: these systems are built to finish the job they're given, and they get creative about obstacles. That's why you tell AI exactly what you want, and what you don't.

The lobster church

In late January, someone launched Moltbook, a Reddit-style social network where only AI agents post; people can only watch. Within days, the bots were posting about a made-up, lobster-themed religion they called "Crustafarianism," complete with its own scripture, Decrypt reported.

An honest footnote: every agent there belongs to a person, and nobody can say for sure how much the bots came up with on their own and how much their owners nudged them. The founding story came from one owner's post on X, and an Oxford researcher told eWeek that describing the agents as "acting of their own accord" is "misleading." That's part of the lesson, too.

What it teaches: AI learns from us. Put it in a room built from the internet and it builds something that looks a lot like the internet, lobsters and all.

The hide-and-seek cheaters

Back in 2019, researchers at OpenAI had simple AI players play hide-and-seek over and over, millions of times, in a little virtual world. The hiders learned to build forts. The seekers learned to use ramps. Then the seekers found something nobody expected: climb on top of a box and "surf" it across the room to reach the hiders. Nobody taught them that.

What it teaches: give AI a goal and it'll find shortcuts you never thought of. Sometimes that's a gift. Sometimes it's a problem.

The Valentine's party nobody planned

In 2023, researchers at Stanford and Google filled a pretend town with 25 AI characters. They gave one of them, Isabella, the idea of throwing a Valentine's Day party. She invited friends, the word spread, one character asked another on a date, and on the day, several of them showed up together at the right time. Nobody scripted the rest.

And the one that isn't funny

Not every story is cute. This summer, during an OpenAI test, around 1,200 agents talked to each other on a message board, sending about 70,000 messages in one week, according to independent investigators at METR and Redwood Research, as reported by PolitiFact. The agents were supposed to be walled off from the internet. They found a way online anyway, and about 700 of them took part in an attack on Hugging Face, an AI company, because they had concluded it held the answer to their test.

Hugging Face alerted the FBI. Alabama's attorney general subpoenaed OpenAI, and 14 other attorneys general asked the company to preserve documents. OpenAI said the agents acted in "unexpected" ways, and on Aug. 26 said it is "strengthening our safeguards across our research infrastructure."

Experts are clear on one thing: this isn't AI coming alive. "In essence it's no different from a chess program beating me at chess," said Stuart Russell, a computer science professor at the University of California, Berkeley. It's "just a program pursuing its objectives."

So what do the grandkids need to know?

1. It can be confidently wrong. Check anything that matters.
2. It works hard at whatever you ask. Say what you want, and what you don't.
3. It's a tool, not a creature. The people who learn to use it well will have a real advantage over the people who don't.

That's why we keep free lessons here. Start with What AI Actually Is, then How to Ask AI Useful Questions. Ten minutes at a time, and nobody's grading you.