Safety mechanisms built into a model to refuse harmful requests or prevent it from generating unsafe content.