How do you handle AI coding tools confidently modifying your existing database schemas?

I had an incident where an agent was supposed to add a simple validation check to an API endpoint, but it decided to improve the database models and added unrequested migrations. It broke half our foreign key relationships because it optimized for what it thought was cleaner architecture.

For teams relying heavily on AI assistants for backend work how do you enforce strict guardrails so they don't touch your data layer without explicit permission?

What strategies do you use to keep AI tools from rewriting database logic you didn't ask them to touch?