Refusal in LLMs is mediated by a single direction — AI Alignment Forum