So we have some internal command line tools at work we want AI to be able to run (on desktop)
But some commands are dangerous so got AI to write a wrapper to discourage the dangerous commands from being run
But the code-scanning security AI says wrapper isn't watertight (we know)