Attack Breaks Permutation-Based Private Third-Party Inference Schemes for LLMs