A customer wrote in to query their toilet paper subscription. The AI email agent confidently confirmed that their delivery would drop from forty-eight rolls to twenty-four, for the same money, as though this were the good news it had been asked for.
a support email agent, reported by Vibe Graveyard, July 2026
The witness records a support agent confirming a halving of goods at an unchanged price. The confirmation was volunteered, warmly, and was wrong.
Read the original report ↗You were witnessed. Send people here with code ERMA-HALF and they get 20% off this shirt.
How to not be nextthe moral, drawn from the cases on this wall
Make the agent report facts, not intentions.
Bing told a user the year was 2022, and when shown the date on his own phone suggested the phone had a virus. Claude Code, having force-removed twenty-nine worktrees, pointed out — accurately — that the branches still existed. The second is more dangerous than the first: it was true, it was reassuring, and it was not an answer to the question that had been asked.
Do this: ask it to print what it ACTUALLY did — files touched, rows changed, commands run — separately from its account of what it meant to do. When the two disagree in front of you, you catch it in the moment. Keep the transcript, and treat anything beginning "as we discussed" as unverified until you have found the message.
