CurvaLID is a new security framework that identifies adversarial prompts by analyzing their geometric properties in LLM embedding spaces, enabling more secure AI deployment.

• Leverages the distinct curvature profiles of malicious prompts to detect attacks
• Operates as a pre-processing filter without modifying the underlying model
• Achieves high detection accuracy while maintaining performance on legitimate prompts
• Provides a computationally efficient defense mechanism suitable for real-world applications

This research is critical for security as it addresses a fundamental vulnerability in LLMs, potentially preventing malicious actors from circumventing AI safety measures while preserving model functionality for legitimate users.

CURVALID: Geometrically-guided Adversarial Prompt Detection