Unvalidated Trust: Cross-Stage Vulnerabilities in Large Language Model Architectures