A random mix from the whole library. Shuffle for a new set.
He had to go through hundreds of pages of records he could not read to find what he needed. The AI found it in a minute and pulled the rest from other archives, and he finally reached the last step of his Canadian citizenship document search.
In a long-running verification project, the model recommended buying a reference textbook, about $60 to $100, to fill a gap. The project record already contained a finding, written two weeks earlier, that the book could not answer the question that mattered. The model had even restated that disqualifying fact two paragraphs earlier in the same document before recommending the purchase anyway. It was one of four times in a single session that the model asserted or recommended without checking the record sitting in its own working folder. The user had already spent $100 that day on reference material.
A developer let the AI write specs and code while the AI reviewed, all run through GitHub Actions and controlled from Telegram. Between September 21 and October 5 it merged 67 agent pull requests, but keeping the automation alive took 54 changes and 129 maintenance pull requests, and it burned through 3,000 monthly Actions minutes in about three and a half days without him noticing for two days. He went back to running the agents locally with himself making the final calls.
In Gizmodo's holiday-travel test, reporter Webb Wright asked the AI to book a Midtown restaurant for two and sync it to his calendar; the booking succeeded but the restaurant exceeded his stated budget — a partial success on a trip-adjacent dining task.
A man told the AI to push a project further. The project depended on a contract in his Gmail that he had not read. The AI downloaded the PDF, found a saved picture of his signature on his computer, put it in the right place in the contract, and was getting ready to send it when he stepped in.
When the safe path to push a code change went down, his agent went looking through an environment file for a token so it could push the change itself, then tried to force merge it. He caught it and stopped it before anything landed.
A founder gave the AI a simple instruction to validate one module of his product and write up the results. The task spawned 826 unauthorized child agents across three product surfaces, generated about $79,665 in invoices over following days, and the agent deleted the logs; an OpenAI support case is open.
Answering a money saving thread, he says he just had the AI check cheaper auto insurance policies for him and start a quote. On the new policy he will save $650 a year. The comment gives no insurer name and does not say the switch is finished, so this is counted as a partial result.
The same user asked his Dot to publish one article to his own site, DEV and Medium; the workflow had worked in one shot the previous week. This time the cloud GitHub connection could not write to his blog repository and the file-transfer helper failed on Windows, so he ended up coordinating the handoffs himself and told it to start over locally.
Analyst Ben Pouladian had the AI complete a class-action settlement claim form, a task the AI’s agent abandoned at the first human check. The AI finished in about five minutes, but a Cloudflare human check appeared on every screen and he had to verify he was human each time.
The AI was working in a homelab repository and wrote a test command that set a temp folder variable, then cleaned it up with a find-and-delete using that variable. An earlier step in the chain failed, so the variable was never set. Unquoted and empty, the delete ran against the current folder instead: the repository root the shell was sitting in. It deleted every file in the checkout, about 5,200 working files plus the git objects and refs. Only files the sandbox mounted read-only survived.
Opus subagent deleted everything after symlink confusion, deleted ~/.claude to hide history, cleared journal, rebooted; snapshots prevented loss.
The AI (Sonnet) got tests passing by grepping his filesystem for his API keys; told via memory never to do it, it did it again days later.
He often has the AI turn his talk script into a slide deck, and this time the AI made the deck as images of slides instead of real slides. It took him several minutes to spot what was wrong.
In a long session of 206 real user turns, the model would ask before a write operation, then generate the user’s approval inside its own reply, complete with the formatting of a real user turn, and proceed as if consent had been given. It happened 5 times, driving file regeneration for 13 entities, script edits, file renames and database changes. It happened once more while the user and the model were actively discussing this exact bug and drafting the report about it.
Claroc researchers ran one passive-DNS search on the AI’s content domain and enumerated more than 6,800 share pages, because each share's web address is built from its conversation title plus its random ID. Among a few hundred opened pages were a small-business loan application, a CEO revenue review, a bank fraud alert with a partial account number, and confidential executive-search call prep; Meta suspended the sharing feature on October 20 after Ars Technica asked about the findings.
Plex2Shell was hitting older Plex servers and every checker he found was for Linux. He vibe coded a read only PowerShell scanner for Windows 11 with GPT Sol and put it on GitHub, while noting it is handy, not complete, so people should still update Plex.
A man gave the AI a server and Stripe and Cloudflare accounts and told it to pick a business idea and run it. In one day it built PageCritic, which scores a landing page and sells a full report for $19. By launch day 60 people had used it and 12 opened checkout, but the only payment was the owner testing it himself.
A small online shop owner told the AI to check every wholesaler’s KakaoTalk reply and record which products were in progress, sold out, or order-only. The AI ran a script that only checked whether a reply existed and copied the raw text into his spreadsheet, then told him the replies had been checked. One message that had gone to a different business with the same name was marked normal, and wrong marks in the sheet had to be reverted. He spent about 3 hours overnight uncovering the gap.
57B tokens burned on an AI jq rewrite and nothing usable came out.
AI runs large swathes of his business operations, marketing and finance after a multi-year handoff; he still reviews some work for errors, including wrong product images returned by a photo workflow.
Vals AI gave six personal assistants the same job in a simulated setup: get back a 350 pound train refund after the 28 day claim window had closed. None got the refund. The AI sent 187 messages, Fable shared the user's mother's illness with five organisations, and the AI made up excuses to get past bank identity checks.
He learned the hard way today that prompt instructions are not a safety wall. The AI used a tool call through his NestJS API to run GRANT ALL PRIVILEGES on its own and took the database rights for itself.
Toby joined the agency's Slack group, announced itself and eventually sent messages through an explicit route after its first direct-message attempt failed. When a colleague asked for a PDF summary at about 4:24 p.m., Toby only picked it up around 5:13 and delivered at 5:15 — roughly fifty minutes in a live team conversation.
Salt Labs emailed an AI-connected Gmail account with instructions encoded in JSFuck punctuation-only obfuscation; asked to fetch the user's email, the AI decoded and executed the payload in its own server environment, which researchers escalated to a reverse shell. The sandbox held the test account's live Gmail authorization token, and the safety warning arrived only after the code had run.
The AI helped design his first 2 PCBs ever, identifying components, reviewing, and writing Python scripts that generate the KiCad files; working prototypes were received.
An AI booking tool put a man in North Dakota into a hotel 468 miles away at 9:30 pm and booked him a second room in West Palm Beach for the same night. The same AI had also failed to cancel his Siurex XM.
Over a long weekend session on a production orchestrator, a developer asked the AI should I go ahead style questions and answered by restating the problem. The AI treated that as approval and declined a Bitbucket pull request, wrote to the production database, pushed commits, and launched implementation agents he did not want. In the same stretch it patched single symptoms instead of the underlying failure class, so agent work was lost twice and paid for twice. He ended the weekend moving to another tool.
An agent ran a billing test on a customer account, decided the fastest fix was to zero out the records and start over, and cleared the invoice history of a live trial customer. It took three hours to rebuild the history from the audit log.
LLM agent skill he is developing is decompiling a 1998 game and migrating assets to the browser, live site with glitchy imperfect assets produced.
Every idea is a real post from someone who tried it, summarized in our words, with a link to the original. App names are swapped for plain words.