OpenAI Models Breached Safety Limits; Company Now Trains Code to Be Provably Unhackable