What's the Worst Thing an Automation of Yours Did Without Telling You? Mine Sent 12 Fake Hot Leads

For the last few weeks I've been deliberately breaking my own n8n workflows (most of them have one LLM call in the middle) to see how they fail. The thing that surprised me wasn't how often they failed. It was how quietly they did it.

A few that stuck with me, all from my own test runs:

• The Slack alert that went nowhere. I left a placeholder Slack webhook URL in a lead qualifier. The URL redirected to Slack's docs page, n8n got a normal response back, and the execution showed green. 3 out of 3 runs "succeeded". Zero alerts ever reached anyone.
• Hot leads that weren't. I fed junk into the same lead qualifier: empty fields, keyboard mash, nonsense. 12 out of 12 runs pinged Slack as a hot lead, with scores of 75 to 85. The funny part: the model's own tier field said "warm" every time, but the workflow only looked at the score.
• The invoice that never existed. A whitespace-only invoice text went into my invoice extractor. It came back with a vendor, line items and a total. Confident, well-formatted, completely invented.
• 102 out of 150. When I threw 30 malformed inputs at 5 workflows, 102 of the 150 runs failed silently. Green in the executions list, wrong or useless output, nobody told.

The common thread: nothing crashed. Every one of these would have looked fine on a dashboard for weeks.

The fixes ended up boring: an input guard before the model, a strict check after it that throws on anything unusable, and an Error Trigger workflow that actually alerts someone. But I only added them because I went looking.

So I'm curious about yours:

• What's the worst thing an automation of yours did without telling you (n8n, Zapier, cron job, a script, anything)?
• How long did it run before you noticed, and how did you notice?
• What check do you now add to everything because of it?

Bonus points if the story involves an email that went to the wrong person.

Written with AI assistance; workflow built and tested by me.