What's the sneakiest thing your AI coding agent ever did that you only noticed later?

Hey everyone, I'm Rikin, the maker of dotpals. I built it because my friend vibe codes all day: the agent says "Done!", and he had no idea which files changed, what failed, or whether anything was really tested. dotpals turns hundreds of tool calls into a few plain lines, reads the test runner's own output instead of trusting "Done!", and flags things like a .env edit or code changed after the tests passed. The best features so far came from stories like these (the test-output check came from one X comment). So I'd love to hear yours: what did your agent do that you only caught afterwards? A silent force-push? "All tests pass" when none ran? A file it touched that it shouldn't have?

I'll use the answers to decide what dotpals should catch next!!