This research introduces the D-SEC threat model, a dynamic approach to LLM security that optimizes both protection against attacks and user experience.

• Addresses the overlooked dynamic nature of adversarial behavior in LLM security
• Balances security measures with usability for legitimate users
• Proposes adaptive defenses that learn from interactions with potential attackers
• Demonstrates improved security-utility tradeoffs compared to static defense methods

This work is significant for security professionals as it provides a practical framework for protecting LLM applications without creating excessive friction for ordinary users, potentially transforming how we approach AI safety in production systems.

Gandalf the Red: Adaptive Security for LLMs