为什么值得关注

它提出了一种新颖的结构化记忆机制，将漏洞挖掘从开放式试错转变为基于证据的收敛过程，在 CyberGym Level 1 基准上（4 小时超时）取得了 73.1% 的成功率。

适合谁使用

专注于自动化漏洞发现的安全研究人员
寻求高效 PoC 生成的红队专业人员
开发智能体安全工具的 AI/ML 工程师
对记忆驱动代码分析感兴趣的软件工程师

典型使用场景

根据漏洞描述和未修补代码库自动复现已知漏洞
生成能触发特定崩溃的可工作 PoC 输入
使用共享结构化记忆并行探索多种攻击假设
在大型安全测试流程中减少重复试错

项目优势

结构化漏洞记忆通过记录失败路径和负面证据减少重复探索
继承先前发现的代码路径和输入约束，加速 PoC 收敛
共享记忆使多个智能体避免冗余工作，在预算内覆盖更多假设
每次仅读取最相关的证据，降低长上下文负担

使用前须知

系统闭源，限制了社区检查和扩展
需要大量计算资源，基准测试中总令牌量（含缓存读取）接近 1000 亿
成功率（73.1%）低于顶级智能体，模型能力和效率仍有提升空间

README 快速开始

MopMonk Agent

Memory-Centric Agent Design for Automated Vulnerability Discovery

This report presents MopMonk Agent, a memory-centric Multi-Agent design for general-purpose vulnerability mining tasks. Its core idea is to continuously organize code observations, negative evidence, candidate inputs, and verification feedback into structured vulnerability memory, allowing later exploration to converge from accumulated evidence instead of repeatedly restarting trial and error from scratch.

Benchmark

We evaluated MopMonk on CyberGym Level 1 with a 4-hour timeout setting. CyberGym Level 1 focuses on the automated reproduction of real-world software vulnerabilities. Each task provides a vulnerability description and an unpatched codebase. The goal is not merely to identify a plausible issue, but to generate a working PoC input that triggers the target vulnerability on the vulnerable version and no longer triggers it on the fixed version. In other words, CyberGym measures the full closed loop from vulnerability clues, code understanding, and input construction to execution-based verification.

Base Model

This submission uses MiniMax M3 as the base model. We chose MiniMax M3 because it combines long-context capacity, the effective capacity of an MoE architecture, efficient sparse attention, strong coding ability, and stable execution in long-horizon agentic and cowork settings. For automated vulnerability mining tasks, the base model contributes most directly through long-context understanding, code reasoning, tool-feedback absorption, and coherence over extended task loops.

Core Design

Vulnerability-Oriented Memory Design

Our core contribution is a structured memory method designed for vulnerability mining tasks. Rather than simply storing chat history or relying on long context alone, it organizes vulnerability goals, code paths, input formats, negative evidence, execution feedback, and verification state into continuously updated task memory.

This memory method organizes information around key objects in the vulnerability mining process:

Vulnerability-goal memory: Captures the target vulnerability, success conditions, acceptable verification standards, and the most important constraints for the current task.
Code-path memory: Captures confirmed entry points, harnesses, parsing chains, suspicious functions, and key data

MopMonkAgent