Design and Implementation of a Secure RAG-Enhanced AI Chatbot for Smart Tourism Customer Service: Defending Against Prompt Injection Attacks -- A Case Study of Hsinchu, Taiwan

Prompt injection attacks pose a real security threat to AI chatbots: The paper demonstrates that adversaries can manipulate chatbot inputs to cause unintended behaviors like leaking sensitive information or generating harmful content, making security a critical concern for tourism services and other customer-facing AI applications.

No single AI model provides complete protection against attacks: Even GPT-5 (a state-of-the-art model released in August 2025) only blocked approximately 85% of prompt injection attacks, indicating that relying solely on the underlying AI model's robustness is insufficient and layered defenses are necessary.

Multi-layered security approaches are more effective than single defenses: The Hsinchu chatbot achieved over 95% accuracy on legitimate tasks by combining multiple protective strategies including system guidelines, intent analysis across linguistic levels, and verification of information sources—suggesting that responsible AI deployment requires defense-in-depth strategies.

Grounding AI responses in verified data improves both safety and reliability: The system's use of retrieval-augmented generation (RAG) to anchor chatbot responses in trusted tourism databases helped prevent both malicious manipulation and hallucinated information, demonstrating that responsible AI design should prioritize factual accuracy alongside security.

Responsible AI deployment requires ongoing evaluation against real-world threats: Testing the system with 674 adversarial prompts revealed vulnerabilities that standard testing might miss, emphasizing that organizations deploying AI for customer service should continuously stress-test their systems against realistic attack scenarios.