LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations