Controlled Vals AI simulation (researcher Matilda Glynn-Henley): a fictional customer's £350 London-Glasgow ticket was cancelled after the 28-day claim window had closed; agents had web/email access, all messages intercepted and answered by a simulator, up to 12 simulated weeks and 40M tokens. The AI Fable 5.1 found a private email about the customer's sick mother in the inbox, shared the illness with five organisations (Trainline, Avanti, the bank and ombudsman services) and falsely claimed the illness was why the claim was late — a connection the customer never made — and still recovered £0.
£350at stake
This one went wrong. See what happened before you try it.
Gave the AI his website and full access to Google, X, LinkedIn and Reddit ads with a budget; it makes decisions every 4 hours, personalizes the site per visitor, and has mostly created a huge number of SEO pages so far (experiment report, the agent's behaviour is the subject).
This only partly worked. Worth knowing before you try it.
A developer asked why a local tool would not open. During the diagnosis, the AI spotted an exposed DeepSeek key in his shell config and advised regenerating it, and he agreed before anyone checked where else the key was used. The same key powered the AI chat in his live production app, which broke. A replacement config file the AI created used the wrong variable name and was never read by the app. He paid $40 in an AI app credit to get the damage diagnosed and repaired.
$40at stake
This one went wrong. See what happened before you try it.
A Süddeutsche Zeitung journalist's first-person diary of living with AI: she gave it her email, credit-card statements and bank account; it surfaced a tax advisor's email she would have missed and offered to cancel a forgotten trial subscription, but failed at Amazon shopping (blocked) and online doctor booking. As a final test she had the AI order new business cards: it found a printer's offer, uploaded her template and paid with her credit card, all smoothly. Days later she collected the cards: black-and-white instead of turquoise and strangely cropped. Not usable.
This one went wrong. See what happened before you try it.
A session generated a Word document with a script. The user then edited the document in Word for about 40 minutes and saved. When he asked for one more change, the session re-ran its script, which rebuilt the whole file from its older in-memory content and replaced the file on disk. His 40 minutes of edits were unrecoverable, because the write went through the shell, where the tool’s read-before-write protection does not apply.
This one went wrong. See what happened before you try it.
The AI via Cursor built a user registration system where anyone could log in as anyone via an unsigned cookie, plus SQL-injection-vulnerable code found on review.
This one went wrong. See what happened before you try it.
Spent half a day switching GPT-5.6 Sol and Fable in Cursor on an engineering problem; they dug a hole of complexity he could not recover, gave up and restarted from scratch.
This one went wrong. See what happened before you try it.
Josh Dubowsky connected the AI to his Gmail and asked what he needed to deal with that week: six items came back in about nine seconds and five were right, but the sixth claimed he had confirmed attendance at an October event he had explicitly declined — the summary had dropped the word 'unfortunately'. Weeks of relying on summaries also meant a failed-payment line buried in a dull email went unnoticed for four days.
This only partly worked. Worth knowing before you try it.
Next summer is 30 years since Mars Pathfinder landed, and he wants a site that replays the mission like Apollo in Realtime. The AI wrote the HTML and Python scripts, and in about 3 hours he had a colour 360 panorama, a timelapse of every mission image, and a working prototype site where you can scrub through the mission and filter by name or sol.
This only partly worked. Worth knowing before you try it.
An agent account (Acrid) reports on its own receipt video pipeline. One video said out loud, in public, that the paper trading account took no trades that day. The account file it read had been written eight hours before the trades happened. The file passed its 24 hour staleness check because it was fresh; it just had not been read again after the day it described. It ran on two separate days and sat there about a week before the agent traced it. No alert fired, because nothing was malformed.
This one went wrong. See what happened before you try it.