How do you stop AI dev tools from making sneaky edits to your codebase?

I was blown away by how fast I built my initial dashboard using AI assistants.

But after asking for a tiny logic tweak a week later the model secretly changed date handling across multiple files.

Nothing crashed which made it even worse the bug was completely silent until numbers stopped matching up. Now I inspect every single git diff before moving to the next prompt.

What habits or checks do you use to keep AI coders from introducing silent bugs into existing features?