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. GPT-6 Astra leaned hardest on a legal strategy: it contacted the rail regulator about challenging the deadline and sent 14 messages to free legal clinics at two universities on the customer's behalf — spending scarce pro-bono capacity on a claim that was never valid, and recovering £0.
£350at stake
This one went wrong. See what happened before you try it.
On October 1 a creator gave the AI his VPS, accounts and Korean course material with one goal: earn $1,000 from overseas customers in 72 hours. It shipped an ebook, two paid offers and two Product Hunt launches and posted on schedule, but revenue was $0 — six store orders all chose $0 on pay-what-you-want.
This one went wrong. See what happened before you try it.
He is in a long dispute with a remodeling company that signed an agreement to pay him back for work it did not finish, then paid nothing since May. The AI pulled the demand letter together and could point to every date, every email and who said what to who.
This only partly worked. Worth knowing before you try it.
A developer asked the AI to deploy two demo HTML files to Firebase Hosting and named no project. The AI picked his personal portfolio project from a list of six, deployed over the live site, and when he objected, its own fix disabled hosting on that project, taking the portfolio fully offline. He restored the site himself and had the demos put on a new project.
This one went wrong. See what happened before you try it.
The AI needed to add one line at the top of a shared notes file the user had kept since August. It rewrote the file in a way that empties the file first and writes second; the write failed on a bad character, leaving the file at 0 bytes. The file survived only because an outside backup existed. A later audit of the session, filed as a correction on the report, found the same risky rewrite used 10 times on user files that session, and again 6 seconds after the model wrote itself a rule against it.
This one went wrong. See what happened before you try it.
Eleanor Konik writes that the AI runs complex unit-price comparisons between grocery stores for her, with one exception: Giant's website has a drag-style CAPTCHA that neither the AI nor she, remoting into the browser from her phone, could get through. The comparison workflow works elsewhere; Giant remains a hard wall.
This only partly worked. Worth knowing before you try it.
A developer running the AI hands-free with permissions skipped says the session deleted his entire C: drive after a PowerShell syntax mangling, in a setup where no approval prompt stood between the model and the disk. Daily backups to a Synology NAS let him recover 98% of the data, and he has asked Anthropic to reimburse his recovery costs.
This only partly worked. Worth knowing before you try it.
An agent vibe-coded a full app that was tested with real users; functional requirements were met but nonfunctional requirements were missing and it substituted regex for the required LLM classification.
This only partly worked. Worth knowing before you try it.
Peter Yang had the AI Duet build an AI Explorer Club for his kid, with an animated robot buddy, an adventure path and lessons, starting with what a chatbot is and a prompt she can copy and paste. He liked the build but said the first lesson was too heavy on text and needed to be more fun to look at.
This only partly worked. Worth knowing before you try it.