GPT-5 introduces a new safety-training method called safe-completion, which aims to balance helpfulness and safety by providing informative responses to dual-use prompts. This approach replaces the previous refusal-based training, allowing the model to navigate complex situations more effectively while adhering to safety guidelines.