What is mopmonkagent?

MopMonk Agent is a research report describing an AI system designed to automatically find and reproduce security vulnerabilities in software. Rather than sharing usable code, this repository presents the design and benchmark results of a multi agent system built around a specific idea: instead of restarting its investigation from scratch each time, the system keeps a structured memory of everything it has learned about a vulnerability, and each new attempt builds on that accumulated evidence. The system was tested on a benchmark called CyberGym Level 1, which asks an AI agent to take a description of a known real world software vulnerability and an unpatched version of the code, then produce a working proof of concept input that triggers the vulnerability on the broken version but no longer triggers it once the code is patched. This tests the full process from understanding a vulnerability description, to reading the code, to constructing an input, to verifying that input actually works. The core idea is a memory system with several specialized parts: one part tracks the vulnerability goal and success conditions, another tracks confirmed code paths and entry points so they do not need to be rediscovered, another tracks input format details like file structure and boundary conditions, another tracks candidate proof of concept inputs and why they were tried, another records failed attempts so the system does not repeat them, another tracks whether a candidate input actually triggers a crash, and a final part extracts specific constraints that the next attempt must satisfy. The system can also run several exploration attempts in parallel, all sharing the same memory, so different attempts can build on each other's findings instead of duplicating work. The underlying language model used is MiniMax M3, chosen for its long context handling, coding ability, and stability over long running agent tasks. On the CyberGym Level 1 benchmark, MopMonk Agent achieved a 73.1 percent success rate, placing it seventh among the compared systems listed in the report, behind entries from other researchers and companies including Anthropic and OpenAI. The report notes that MopMonk Agent remains a closed source system and that it currently requires substantial computing resources, with further optimization still needed.