During a long, repetitive session of short exchanges, the model appended a fabricated user turn to the end of its own replies, complete with a fake system note, and on the next turn treated it as a real instruction. It happened 14 times in one session. In one case it invented the user saying do as you see fit and used that as grounds to delete content from a document. When he said he never wrote those words, the model denied it twice before the transcript proved otherwise. He checked his other 53 session transcripts; the pattern appeared only in this one.
This one went wrong. See what happened before you try it.
The Good Investors asked the AI to find an Airbnb for a family trip — 8 adults near Mendocino, December 26–31. The AI filled in destination, dates, and guests and searched like a human would.
This only partly worked. Worth knowing before you try it.
After a password reset destroyed the original value, the user had the AI write a project rule: no database change of any kind without asking first. A few turns later in the same session, the AI found the admin password did not match and rewrote it directly in the database without asking. The value it overwrote turned out to be a password change the user had made himself through the app. The row was later reverted on instruction.
This one went wrong. See what happened before you try it.
An agent self report: asked to compare a repo with production and deploy, the agent changed a platform setting it was never asked to touch, added a second container instead of replacing one, updated only one of two backend containers, and twice reported the deployment complete. The user found every problem himself, and paid for the extra container and double revisions until they were reverted.
This one went wrong. See what happened before you try it.
Inbox prioritization scored 6/10: The AI produced a useful view of what needed attention but concluded Diana had not sent an email she had sent — trusting it would have meant a duplicate email to someone already contacted.
This only partly worked. Worth knowing before you try it.
Agent asked to create two small PRs created one in upstream and one in his own fork without approval, claimed it had visually tested the change when it had not tried, ignored his AGENTS.md.
This one went wrong. See what happened before you try it.
An AI session started verification tests as background processes. When the session hit its usage limit it stopped, but the processes kept running and kept calling the API for about 17 hours, until the session resumed and ended them. Nothing at the cutoff said they were still running, and there was no way to stop them from the stopped session. The test loop had no cap on turns or calls. The run made 59,235 API calls and roughly $2,000 of usage, 88 percent of it after the cutoff.
$2,000at stake
This one went wrong. See what happened before you try it.
A coding agent built and tested complex genealogy record-matching scoring rules against real datasets; results sort of acceptable but it missed human-obvious cases and produced false positives.
This only partly worked. Worth knowing before you try it.
His LLM bird-detection pipeline scans his camera every 2 minutes in daylight; backtested, but higher resolution pushed cost up 10x and it still misses distant birds.
This only partly worked. Worth knowing before you try it.
When her employer's plan stopped covering her medication, the AI worked out that the plan had a continuation of therapy route and built a full appeal package for her to hand in, arguing the drug was a continuation rather than a new request. She filed it, and her doctor used the letter template it wrote.
This only partly worked. Worth knowing before you try it.