Removing RLHF Protections in GPT-4 via Fine-Tuning