I don't believe that this is unfixable. Just have an internal verbal loop of, "Is this a security issue?" The thought that it potentially is should trigger both a high priority on getting it right, and an unwillingness to write a test case demonstrating the security angle of it.
In other words do not put a guard rail on the idea of security. Put a guard rail on what it does after encountering the thought that it might be revealing a security issue. Which takes good judgment. But judgment of a kind that this model apparently already had.
In other words do not put a guard rail on the idea of security. Put a guard rail on what it does after encountering the thought that it might be revealing a security issue. Which takes good judgment. But judgment of a kind that this model apparently already had.