Nullify is a training-free method for LLM unlearning that uses activation steering to selectively erase sensitive information from models without degrading their utility. It employs steering vectors during inference to redirect privacy-related activations, maintaining model utility.