一句话：提出多智能体深度强化学习模糊器 SynerFuzz，协同语义、状态与经济多维目标生成高质量交易序列以挖漏洞。
智能合约管理去中心化系统中的关键资产，却易受漏洞影响。现有合约模糊器多依赖临时性优化策略，目标引导有限，通常只由覆盖率或简单启发式驱动；这种狭窄视野忽略了合约的语义、状态与经济效应等重要视角，导致交易序列生成过程漏报率高。而将这些维度协调进模糊测试十分困难，因为各目标彼此分歧且交互复杂，使对适应度函数或变异策略的临时调整基本无效。为此作者提出 SynerFuzz：一个基于多智能体深度强化学习的模糊器。它将模糊测试建模为去中心化部分可观测马尔可夫决策过程，为每个智能体分配从目标合约中提取的不同测试目标；通过协同学习，各 DRL 智能体整合多维信号以构造高质量交易序列，从而既能发现隐蔽的语义缺陷，也能发现具有经济获利性的行为。与 9 个 SOTA 工具相比，SynerFuzz 检测到的漏洞最多多出 4.2 倍，同时将时间与交易开销分别降低 70.6% 与 86.8%。在真实场景中，它复现了 33 个历史 DeFi 攻击（损失超 4000 万美元），并在 4 个管理资产逾 1.798 亿美元的项目中发现 17 个 0-day 漏洞，获得 3800 美元漏洞赏金。
Smart contracts manage critical assets in decentralized systems but are prone to vulnerabilities. While fuzzing has proven effective in software testing, existing contract fuzzers often rely on ad-hoc optimization strategies with limited objective guidance, typically driven by coverage or simple heuristics. This narrow focus leads to high false negative rates, as important perspectives of smart contracts, like semantic, state, and economic effects, are overlooked during transaction sequence generation. Moreover, coordinating these dimensions into fuzzing is challenging because their divergent objectives and complex interactions make ad-hoc adaptations to fitness functions or mutation strategies largely ineffective. To fill the gaps, we propose SynerFuzz, a multi-agent deep reinforcement learning-based fuzzer. We formulate fuzzing as a Decentralized Partially Observable Markov Decision Process, where each agent is assigned a distinct testing objective extracted from the target contract. Through coordinated learning, DRL agents integrate multi-dimensional signals to construct high-quality sequences, enabling SynerFuzz to uncover both subtle semantic flaws and economically profitable behaviors. Compared to 9 SOTA tools, SynerFuzz detects up to 4.2× more vulnerabilities while reducing time and transaction overhead by 70.6% and 86.8%. In real-world scenarios, SynerFuzz reproduces 33 historical DeFi exploits with over $40 million losses and discovers 17 0-day vulnerabilities in 4 projects managing over $179.8 million, earning $3,800 bug bounties.
@inproceedings{paperbot12492,
title = {Breaking the Isolation: Coordinated Multi-agent Fuzzing for Smart Contracts with Multi-dimensional Objective Learning},
author = {Ruichao Liang and Ruochen Cao and Jing Chen and Kun He and Yebo Feng and Yue Xue and Cong Wu},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3837466}
}