Understanding the Dark Side of LLMs' Intrinsic Self-Correction