Dynamic Collaborative Intelligent Multi-Agent System with Planner and Heterogeneous Executors for Offensive Security

View the PDF file from the paper entitled D-CIPHER: Multi-Cooperative System System with Planning and Insecure Portes of Security attack, through a Udeshi, Minghao Shao, Haoran XI, Nanda Rani and Kimberly Milner Milner Milner Milner Krishnamorine, Varshu Guardi, Ramish Carrie and Mohamed Shavek

PDF HTML (experimental) view

a summary:LLMS models were used in cybersecurity, such as independent security analysis or penetration test. Capture The Flag (CTF) challenges are criteria for assessing the automatic task planning capabilities of LLM Cyber ​​Security. Early attempts to apply the LLMS application to solve the challenges of the CTF challenges used as the systems of agents, where the comments were restricted to one logical workshop. This approach was not enough for the complex CTF tasks. Inspired by the CTF competitions in the real world, where teams of experts cooperate, we offer the D-Cipher LLM with a CTF cooperative solution. D-Cipher combines the factors with distinct roles with dynamic feedback rings to enhance thinking of complex tasks. The system of the executive agent, which consists of a planning agent to solve comprehensive problems alongside the multiple heterogeneous implementation agents, provides the multiple homing agents for individual tasks, which facilitates allocating effective responsibilities between factors. In addition, D-Cipher includes a car agent to improve problem solving by automatic generation to demand a very relevant initial. We evaluate D-Cipher on multiple CTF standards and LLM models through comprehensive studies to highlight the effect of our improvements. In addition, we manually plan CTFS in the bench at New York University CTF for MITER Att & CK technology that applies to a comprehensive assessment of the offensive security capacity of D-Cipher. The D-Cipher achieves the latest model on three criteria: 22.0 % on the NYU CTF seat, 22.5 % on Cybench, and 44.0 % on hackthebox, which is up to 2.5 % to 8.5 % of the previous work. D-Cipher reports 65 % of ATT & CK technologies compared to previous work, indicating a stronger offensive capacity.