Lab

Guardrails for an AI that writes SQL

If you let a language model query your database, asking it nicely isn't enough. Write a query, or try an attack, and see which layer stops it. Nothing is executed: the database is invented.

Try:

And what comes out, masked

Even when a query passes, personal data never reaches the model: it's hidden by column name and also by content, even inside JSON.

Before


      

      

        
After