一句话:提出XploGen,用业务逻辑的oracle规约引导LLM合成智能合约漏洞利用,兼顾多样性与具体exploit示例。
智能合约漏洞分析工具常在有限数据集上训练和评估,存在三方面缺陷:ground truth 多依赖同类工具的判定而非真实验证结果;数据集集中于重入、溢出等低级漏洞;且缺少具体漏洞利用示例。为此,作者提出 XploGen,用基于模型的 oracle 规约刻画合约业务逻辑,并借助 LLM 合成针对性的有效漏洞利用。实验构建了 104 个合成漏洞-利用对,在合约特定方面的利用成功率达 57%,且每个利用平均仅需 3.5 笔交易,表明该方法能以较高效率生成多样且具体的漏洞利用。
Many smart contracts are prone to exploits, which has given rise to analysis tools that try to detect and fix vulnerabilities. Such analysis tools are often trained and evaluated on limited data sets, which has the following drawbacks: 1. The ground truth is often based on the verdict of related tools rather than an actual verification result; 2. Data sets focus on low-level vulnerabilities like reentrancy and overflow; 3. Data sets lack concrete exploit examples. To address these shortcomings, we introduce XploGen, which uses a model-based oracle specification of the business logic of the smart contracts to synthesize valid exploits using LLMs. Our experiments, involving 104 synthesized vulnerability-exploit pairs, demonstrated a 57% success rate in exploiting targeted aspects of the contract. They achieved exploit efficiency with an average of only 3.5 transactions per exploit, highlighting the effectiveness of our methodology.