Trust Vulnerability Exposed in RAG Architecture: $12 Fraud Breaks the Acceptance Logic of Mainstream Large Language Models