G2FUZZ introduces a cost-effective approach for using LLMs to generate complex non-textual inputs (images, PDFs, videos) for comprehensive software fuzzing.

• Leverages LLMs to synthesize input generators as code rather than directly generating the inputs
• Creates grammar-aware fuzzing capabilities for complex file formats that LLMs struggle with directly
• Achieves competitive bug-finding performance while maintaining lower computational costs
• Provides a practical solution for security testing across diverse input formats

This research significantly enhances software security testing by enabling efficient fuzzing of applications that process non-textual files, addressing a critical gap in LLM-based security tools.

Low-Cost and Comprehensive Non-textual Input Fuzzing with LLM-Synthesized Input Generators