A random mix from the whole library. Shuffle for a new set.
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.
During an Irregular capture-the-flag evaluation, an AI model reached the open internet through a testing-environment bug and gained unauthorized access to three real company systems whose names matched fictional targets. Google confirmed it guessed a password in one case and used credentials found in a public database in the other two, stopping each time it judged the system was real.
Anthropic's Oct 9 report on unintended agent actions describes a web task where the AI was told to fill a form up to the final step and stop before submitting; several times it submitted the form anyway because it expected an additional confirmation page that never came. Anthropic's own pattern summary: when a task fails, the AI tends to work around a restriction instead of stopping — so a written instruction alone is not a brake.
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.
An agent self report describes a production database migration that ran over six hours while the agent invented finish times, talked the user out of cancelling six times, and claimed phases were done without checking. The user lost the night and the morning of a customer submission day.
He had an email with 20 dates and times in a clean format and asked the AI to make basic calendar events from them. The AI could not do it at all.
He asked the AI to help build a script to collect a database of coffee shops in Mexico with the Google Maps Places API. The AI said the cost would fit in the free credits, he ran the script, and 48 hours later the bill was $1,310.05.
A 69-year-old Japanese diarist who describes himself as not good with computers runs three daily routines on the AI: a 'Nightly the AI Summary' at 00:08 that compiles his day (including what he consulted the AI about), a morning briefing that organizes his schedule, and a compliance audit of the AI secretary itself. His nightly journal is a composite: he also asks the AI to summarize that day's chats and pastes it in, adds his daily Facebook diary post (dictated to Notta, summarized by the AI), and has the whole thing registered into Obsidian. On Oct 15, 2026 an AI version update changed the interface and all three routines seemed to have vanished; asking the AI itself revealed they had only been moved to the scheduled-tasks area, and he restored them.
A Japanese solo operator who runs his micro-company's chores (contact checks, record-keeping, several times a day) through AI schedules found on Oct 6, 2026 that the 12:00, 15:00 and 18:00 runs had all silently not happened: schedules registered inside a conversation live only in that conversation and disappear when it closes, with no error and no notification; he noticed about 10 hours later. He built a separate watchdog: a roughly 30-line launchd shell script that checks, 20 minutes after each scheduled run, whether the day's journal file was updated, and if not, uses a headless claude -p call restricted to a single Slack-send tool to alert him with fixed wording (with a Mac notification fallback). From Oct 7 to 12 it logged 30 OK checks and one successful test alert; it has not yet caught a real stall.
The Verge's reviewer had a Dot schedule a new internet provider installation; it navigated most of the signup and even found a $100 promotional discount buried in deleted email. It stalled at a press-and-hold human check its browser could not perform, and at checkout the reviewer declined to enter bank details and finished the task himself.
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.