Abstract
Penetration testing is critical for identifying system vulnerabilities. Traditional manual methods, however, are laborintensive and rely heavily on scarce expert knowledge. Existing automated tools often require complete network visibility or specific pre-trained models, limiting their applicability in blackbox scenarios. To address these limitations, we propose PenAgent, a MetaGPT-based multi-agent framework that automates the entire penetration testing lifecycle. The system coordinates four specialized agents to perform reconnaissance, scanning, exploitation, and logic optimization via Retrieval-Augmented Generation (RAG). We evaluated PenAgent against ten N-day vulnerabilities, demonstrating that GPT-4-based agents achieved an exploitation success rate exceeding 80%. Furthermore, RAG integration significantly boosted the performance of smaller models (e.g., GPT-3.5-turbo) to levels comparable with GPT-4. These findings underscore the potential of collaborative LLM agents in autonomous security assessments.