GradCoo introduces a resource-efficient method to detect unsafe prompts in LLMs by analyzing gradient co-occurrence patterns without extensive data requirements.

• Identifies unsafe prompts by examining consistent patterns in gradients of safety-critical parameters
• Requires only a few reference examples rather than large-scale fine-tuning datasets
• Achieves effective detection with significantly lower computational costs than traditional guardrail approaches
• Offers a practical solution for enhancing LLM security while preserving utility

This research represents an important advancement in LLM security, providing developers with efficient tools to prevent harmful outputs while maintaining model performance in legitimate use cases.

Gradient Co-occurrence Analysis for Detecting Unsafe Prompts in Large Language Models