I would be very surprised to see any "make it safe" solution that does not render these models useless for programming. I assume posts like this are mostly clout-motivated, but I really dislike the implied vision for what LLMs should offer here.