[dkersten]

You could train a model on purely secure code and still have it combine it in insecure ways.

[0-_-0]

And the other way around.

[SketchySeaBeast]

I guess the question really is how many ways are there to do it wrong, and how many ways are there to do it right?