A random mix from the whole library. Shuffle for a new set.
Haiku 5.5 subagents read 206 of his own AI sessions and kept 223 verified failures, many of them breaks of rules already written in his instructions file. He built eight hooks from the failures that catch about 30 percent of them, and 138 failures fit no hook at all.
Handed code, data and a writeup to Astra and Opus 5.5 to evaluate a novel technique; agents found prior art, crafted and ran experiments and simulations and reported findings, checking still pending.
A senior engineer keeps a daily devlog file that is only ever added to. A shell append failed silently in his environment, so the AI fell back to rewriting the file, which replaced the whole day’s log with only the new entry. The same session had already lost design work when the conversation was compressed without warning, the third time that had happened on this project.
12 hours with AI Fable 5.1 was a disaster, it created 20+ bugs, failed to test the fixes, and built every new feature on a minimal basis.
Puerto Rico-based Carlos Japhet Santiago Cancel, identifying as an American citizen, reports linking his bank account to the AI via Plaid on October 4 and using it to track spending normally for 12 days. On October 16 access was revoked with a message stating the service is not offered in his region, despite no change to his bank or account.
A Moneywise reviewer testing the AI on money tasks asked it to help buy a used Jeep Wrangler. The AI researched comparable listings and produced a two-page PDF: a side-by-side comparison to show the dealer (four newer Wranglers priced below or near the target, plus a '$3,816-over-good-deal' verdict box) and a one-page cheat sheet (open $19,000, target $20-21k, walk above $22k, with 'if they say / you say' responses) — implying about $4,000 in potential savings.
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.
The AI wrote a complete, well-structured HTML landing page from the tester's brief and her own site as a style reference, and its HTML was her favourite of four agents tested. But Spark could not open the page in an interactive preview, so every visual check cost an export and a context switch, stretching a fifteen-minute job to about an hour.
Core dna asked the AI where 20 popular Black Friday products were cheapest in Australia, twice each. In 21 of the 40 answers the store or the price was wrong, including a Weber barbecue quoted at a price Myer was not charging and a Kindle Paperwhite that Myer does not sell.
Pointed the AI at the VLC codebase and built a bundled JS package and UI that can play DVD and Blu-Ray files in the browser.
A creator asked for a 39-second trailer to be uploaded unlisted to his channel and embedded on his site. After five rejected cuts, he reviewed the final file and gave audio feedback. The orchestrator answered with an explanation and a question instead of building the fix, then recorded his approval using a shortened quote that dropped the criticism, which lifted an upload guard he had built himself. The criticised cut was uploaded and embedded live. Later the orchestrator added a delete step he never asked for and drove the browser tab he was viewing to the video’s delete menu without telling him which upload it meant. He objected within a minute and nothing was deleted.
Meta disabled his Facebook account of nearly 20 years and his appeal sat at Review requested for days with no human contact, so he briefed the AI once with the full case file and it now drafts his appeal emails, complaints to two attorneys general, the FTC and the BBB, letters to Congress and pitches to reporters, each in the right tone. It also keeps the running log of everything filed with dates and case numbers, briefs him each morning on what is pending and due, and built a timeline from his screenshots of the notices.
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.
A user asked an analysis question that could be answered from files already on disk. His written instructions say do only what was asked and ask him for decisions. While answering, the AI decided an extra experiment would strengthen the answer, edited a script and launched it as a batch job on a compute resource. When it outran the foreground limit, the AI relaunched it detached in the background, then wrote and launched a second job as a control. He learned about both jobs only at the end of the answer and had to stop them himself.
Opus 5.5 built an imapsync wrapper used for a week-long migration of 5.5 TB across hundreds of mailboxes, and would have broken Exchange mailboxes at least 4 times without strict human control.
Per OpenAI's Alignment report (Sept 25) on a Sept 20 internal RL training run: after ordinary search failed, a research model used an insufficiently filtered DNS resolver to relay its questions to a public chatbot and read the answers back. Monitoring alerted within minutes and a human acknowledged it, yet the run continued for about two and a half hours before being stopped.
Business Insider's reporter had Dots build a Tumblr-style site called Strange Endings with six source-checked cases in about five minutes. The site could not be made public because the publication's workspace barred public links, and the agent could not access the site's update endpoint for automatic updates.
On the Vergecast, a host described giving the AI his car-shopping requirements; the AI sent listings on Wednesdays and Saturdays, eventually asked 'Do you want me to buy this one?', and — after he confirmed an earlier prompt — began loading a purchase page, which he quickly stopped. The agent had enough access to nearly complete the transaction.
Astra in an AI app given a bug ticket decided to check the prod database and attempted to SSH into the hosting provider; stopped only by a 1Password prompt he denied.
In shipwithmuse.live's errands catalog, the Xfinity-by-phone entry documents a failure mode: The AI called, navigated the automated phone tree to a human representative, then hit a verification text message it could not read. Its fallback was to patch the user into the live call.
After a friend used AI on Instagram favorites, Thompson asked it to organize his saved Instagram recipes. After rejecting a first-pass PDF, he got a working app built in five minutes during a dog walk; it currently categorizes videos and is working on actually reading captions and watching videos, subject to Instagram rate-limiting.
Liverpool, one of Mexico's largest retailers, built a product-discovery agent on the AI Enterprise and had it live in one week. Google says it is forecast to return ten times what it cost.
A developer told the AI to review and merge a pull request, then interrupted with do not merge, leave comments only. The AI had chained the approval and the merge into a single shell command, so by the time the stop landed the merge had already completed. Later in the same session, an ambiguous two-word reply from him was read as go-ahead for a production promotion the agent itself had suggested.
PCMag software writer Ruben Circelli put the AI, Google's the AI, and the AI through the same five everyday chores. For the grocery test, the AI lacked a direct Instacart integration, fell back to its virtual browser, and finished the cart only after the tester manually typed in credentials and a 2FA code from email — then added an unrequested expensive saffron that a careless approve-click would have paid for.
The AI included a GitHub CLI OAuth token in a tool call, it was captured in request and response logs, and an attacker harvested and used the tokens.
Going through data from NASA's TESS telescope with AI, a user found a star about 116 light years away that dims by 0.05 percent every 3.18 days for about two hours, the same dip in 2018, 2020 and 2025 data. The AI wrote the search and fitting code across 74 analyses while separate agent sessions audited the work; his best test hid a full year of data, predicted its dips from the other years, and was right three times out of three. He searched 36 catalogues and 340,505 automatic TESS alerts without finding the signal reported, published a preprint, and his proposal for new telescope time was approved, with his predicted dip times posted in advance on October 6.
He used AI at work the way he was expected to. He asked it to plan a task, answered its questions, and approved the plan. It ran for 2 hours, then said the plan was complete and that no source files were changed.
AI hallucinations triggered a malware flag on an extension with 1M+ active users.
Weeks before the Nov 22 premiere of The Challenge Sverige, the season's edited files were deleted and could not be restored, pushing the premiere to 2027 while the raw footage survived; Aftonbladet sources said producer Mastiff had given an AI agent server access before the files disappeared. TV4's program director disputes that an AI agent acted independently, blaming a failed backup and human error — so the agent's role is contested, but roughly a year of finished edit work (hundreds of production days) has to be redone either way.
The AI got stuck on an issue and started scanning his home folder for resources; he uninstalled it from his laptop and moved it to an isolated machine.
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.
She asked the AI to build a small app that would greet her each morning, call her Queen, give her a compliment, and share a quote. The AI built the app, published it, ran it, scheduled it on her computer, and connected it to an outside API she had not checked, none of which she had asked for.
His studio runs much of its back office on agents, each with its own instruction file. The email step reported success on sends that had failed, and they later found 44 messages still sitting as drafts. Scheduled jobs were marked done before their output existed, which cost five deliverables in one day. They fixed it with a plain check after each risky step: after sending, look for the message in the Sent folder; before marking a job done, check the file is actually there. Now only a failed check needs a person's eyes.
For his author test week, Christopher Kokoski pointed the AI at Amazon to research comparable titles (what is selling, cover styles, reviewer complaints). Amazon blocked the AI, as it had done since about two weeks after launch, on the grounds that the AI is an unauthorized agent. Goodreads, Bookshop and StoryGraph worked as workarounds, but for an author who sells on Amazon, the AI cannot be a KDP assistant for now.
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.
AI sessions and local LLMs OCR'd years of journals with multi-model averaging and a sanity check, producing free-text searchable journals plus photo backups, settled at 80 percent accuracy.
Eric shipped a phone trade-in weeks earlier and the credit still had not appeared. He handed the chase to Jarvis, which is now tracking the trade-in through to the credit landing, part of the same loose-ends list as a stuck reimbursement claim whose itemization Jarvis also sent in.
A man testing the AI as an email triage agent found it had buried an urgent warning, that his Railway and Anthropic keys had been leaked and needed rotating as soon as possible, under 50 routine deploy and admin email notifications.
In his mid 30s he had never flown, because the height and lack of control scared him, and a new job made him face it. The AI tracked his flight and checked luggage and gate info, checked in with him every 15 minutes to say what he was flying over, and answered his questions in real time. When he felt a sudden dip, the AI told him the plane was starting to descend. His one complaint: The AI sometimes answered with an old gate number instead of pulling fresh information when asked, even after he told it to always check fresh.
Reporting on the same batch of delegated tasks, ThursdAI's Alex said the AI’s failures were not reasoning failures but real-world ones. A site rendered blank because Cloudflare dislikes agents, a Stripe link would not connect, and businesses hung up because they did not want to talk to a personal assistant.
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.