Bridging AI and software security: A comparative vulnerability assessment of LLM agent deployment paradigms