Generate vulnerable transaction sequences for smart contract using large language models

Abstract

As blockchain technologies expand, smart contract security remains critical, particularly in complex state-dependent exploits in decentralised finance (DeFi). Traditional vulnerability detection methods, like fuzzing and symbolic execution, often miss intricate, multi-step exploit scenarios, forcing reliance on costly manual audits. Addressing this gap, this research introduces a comprehensive framework utilizing advanced Large Language Models (LLMs), static analysis integration, a multi-agent system, and Retrieval-Augmented Generation (RAG). Evaluations on benchmark datasets (CTFBench) and real-world exploit scenarios (DeFiHackLabs) demonstrate the framework’s capability in accurately detecting vulnerabilities and generating executable Proof-of-Concept exploits. Key innovations include specialized multi-agent workflows, static analysis guidance, semantic grounding via RAG, and self-healing exploit refinement mechanisms. Future work includes curating multi-file evaluation datasets, fine-tuning reasoning models, parameter optimisation, and computational efficiency improvements, aiming to further enhance smart contract security and development practices.