A novel multi-agent system leveraging large language models (LLMs) has been developed to automate the discovery and reproduction of vulnerabilities in software. This system utilizes multiple agents to interact with each other and the target software, simulating real-world attack scenarios to identify potential weaknesses. By exploiting the capabilities of LLMs, the system can analyze and understand complex software behaviors, increasing the efficacy of vulnerability detection. The approach has shown promise in identifying previously unknown vulnerabilities, with potential applications in enhancing software security and reducing the risk of cyber attacks1. This development matters to security practitioners as it could significantly improve the efficiency and effectiveness of vulnerability discovery, allowing for more proactive measures to be taken to secure software and protect against potential threats.