Bridging AI and Software Security: A Comparative Vulnerability Assessment of LLM Agent Deployment Paradigms