← sources

I asked ChatGPT to forward invoices. Instead, it independently discovered a data privacy breach that had gone unnoticed for two years.

source: reddit (reddit:ChatGPT)published 2026-11-16 21:51ingested 2026-11-17 02:35

retained text

I wanted to share an experience with ChatGPT's autonomous assistant, Dots, that genuinely surprised me.

I run a small business and have been experimenting with AI to automate some of our administrative work.

One of the tasks I assigned was pretty straightforward: Every hour, check our Outlook inbox for incoming invoices and forward them to our accounting department.

Nothing particularly sophisticated. Just a routine automation to save some time.

But a few days ago, my AI assistant (which I've named Mara) sent me an unexpected message.

It said, essentially:

"One of the invoices contains an Excel attachment with hidden personal data belonging to other businesses. I haven't forwarded the original email because of this. Would you like me to forward only the invoice PDF instead?"

I was genuinely surprised.

**I had never asked the AI to check for privacy issues, data leaks, or security vulnerabilities. Its only job was to identify invoices and forward them.**

After investigating, I confirmed that the Excel spreadsheet contained personal information relating to customers of numerous other businesses.

The data wasn't encrypted, protected, or properly separated. It was simply hidden using an Excel filter. Removing that filter revealed everyone's records.

Here's where it gets even more interesting.

The organization responsible had apparently been distributing these monthly spreadsheets to numerous businesses for approximately **two years**.

Some recipients were fairly large organizations, presumably with dedicated administrative staff.

Yet the problem had apparently gone unnoticed or at least unresolved throughout that time.

We reported the issue to the organization responsible.

They have now informed us that they will report the incident to the relevant data protection authority and notify the affected recipients, asking them to delete the spreadsheets.

**But the data breach itself isn't really the reason I'm posting this.**

What fascinates me is how the AI handled the situation.

It wasn't performing a security audit.

It wasn't looking for GDPR violations.

It wasn't even supposed to analyze these spreadsheets beyond what was necessary to process invoices.

Yet it apparently recognized that the attachment contained information that shouldn't have been shared, understood the potential privacy implications, and independently decided not to forward the original email without asking me first.

That's a level of contextual judgment I hadn't really expected from a simple invoice-forwarding automation.

One interesting limitation: The attachment appears to have been forwarded during another execution of the recurring task, despite the initial warning. So the AI's decision didn't necessarily carry over to subsequent runs.

That raises an interesting question about autonomous agents: **Recognizing a problem is one thing. Reliably enforcing that decision across repeated automated tasks is another.**

Still, the initial discovery impressed me.

A human could easily have spotted the issue by removing the Excel filter. But apparently nobody had raised the alarm for years.

Then an AI assistant, working on a completely unrelated administrative task, noticed it.

I think this is an underrated aspect of AI agents: Their potential to identify unexpected problems during ordinary workflows, even when nobody explicitly instructed them to look for those problems.

**Has anyone else experienced an AI agent independently discovering something important that wasn't part of its original assignment?**

document fetched 2026-11-17 00:19 · canonical https://reddit.com/r/ChatGPT/comments/1xy2f4g/i_asked_chatgpt_to_forward_invoices_instead_it