LLM Framework Detects Smart Contract Vulnerabilities with High Recall

A new framework utilizing LLM technology has been developed by researchers to identify security weaknesses in smart contracts on various blockchain platforms. This system employs AST-based context extraction along with tailored prompts for specific vulnerabilities, enabling the creation of customized detectors for 13 common vulnerability types. A comprehensive dataset comprising 31,165 expertly annotated instances of vulnerabilities from more than 3,200 real-world projects across 15 leading blockchain platforms has been assembled and made available. Experimental findings indicate a robust average positive recall of 0.92 and a negative recall of 0.85, showcasing its effectiveness. This framework overcomes the shortcomings of current detection methods that are inflexible regarding vulnerability types and depend on manually created expert guidelines.