Why Are Web AI Agents More Vulnerable Than Standard LLMs?