Tricking LLM-Based NPCs into Spilling Secrets

Large Language Models (LLMs) are increasingly used to generate dynamic dialogue for game NPCs. However, their integration raises new security concerns. In this study, we examine whether adversarial prompt injection can cause LLM-based NPCs to reveal hidden background secrets that are meant to remain undisclosed.

Key Contributions

• Identifies and demonstrates a security vulnerability in LLM-based NPC dialogue systems where prompt injection can extract developer-embedded hidden secrets
• Constructs a fictional game world with embedded NPC secrets to empirically evaluate the susceptibility of LLM-powered NPCs to adversarial prompt injection
• Highlights the under-studied security surface of custom developer-defined NPC system prompts as distinct from standard LLM safety mechanisms

🛡️ Threat Analysis

Details

Similar Papers