What they did
The poster runs a small business and was moving admin work onto a ChatGPT Dot, OpenAI's always-on assistant, which they named Mara. One recurring task ran every hour: look through the company's Outlook inbox for incoming invoices and forward each one to the accounting department. Nothing in the instructions mentioned privacy, security or checking attachments beyond spotting invoices.
How it went
A few days in, the Dot messaged the poster instead of forwarding one email. It said the Excel file attached to that invoice held personal data about other businesses' customers, and asked whether to forward only the invoice PDF. The poster removed an Excel filter and found the records of many recipients, unprotected, in a monthly file the sender had been mailing out for about two years. They reported it, and the sender said it would notify the data protection authority and ask recipients to delete the files. The hold did not stick: a later run of the same hourly task forwarded the attachment anyway.
Worth knowing
A judgement made in one run did not carry over to the next run of the same recurring task, so a flag from the agent needs a rule added to the task itself, not just a reply. Commenters pointed out that this inconsistency means you cannot count on the catch happening again. The poster used AI to help write the post.