I asked my AI coding agent a yes-or-no question.

It made a commit instead.

I never asked it to change anything.

When did "can do" become "authorized to do"?

This incident changed how I think about AI agent security.

Here's what happened →