Typically there's a firewall between the web server and the middle tier. And speaking an application-specific protocol across this boundary does improve security. The hacker is constrained to used the application-specific protocol, and is blocked from using raw SQL, thus limiting the amount of probing they can do.

Eventually the only practical solution comes down to controlling access to the web account, and making the CGI read-only to the web account.

And if you lose the box to some other exploit, the hacker gets a valid database username and login. Even if the pair that the CGI uses is for an INSERT-only account, if the database is on the same box, it's toast.