ANALYSIS OF METHODS FOR DESIGNING THE ARCHITECTURE OF A SECURE MULTI-AGENT SYSTEM FOR AUTONOMOUS PENETRATION TESTING BASED ON LARGE LANGUAGE MODELS

Abstract

The article provides a comprehensive analysis of methods and approaches for designing the architecture of a secure multi-agent system for autonomous penetration testing based on large language models. The relevance of the study is driven by the need to bridge the gap between resource-intensive, slow manual information security audits and the dangerous full automation of classic scanners, the uncontrolled use of which in production environments poses critical risks of Denial of Service (DoS) or exceeding the authorized scope. An analysis of recent publications revealed that existing AI-based solutions (e.g., PentAGI or Penligent) suffer from specific threats such as cascading failures, identity spoofing, and group hallucinations of language models. The theoretical foundation of the work is the concept concept of integrating a neural network approach and cryptographic protection of operations, according to which large language models exclusively perform the functions of semantic context management, strategic planning, and task routing. The direct execution of commands and the generation of evidence are delegated to real specialized tools, eliminating the risk of artificial artifact fabrication. The research methodology involves designing the backend infrastructure as a modular monolith using TypeScript and the Nest.js framework based on CQRS and Saga architectural patterns. To manage the attack deployment process, the LangGraph framework is utilized, formalizing the audit as state graphs with a clear distribution of tasks (reconnaissance, vulnerability search, verification) among independent agents. The main result of the research is the developed mechanism for strict tool isolation: all destructive utilities, including Metasploit, nmap, and OWASP ZAP, are launched in disposable Kali Linux-based Docker containers. Communication between the orchestrator and the isolated environment occurs solely through standard input/output streams according to the Model Context Protocol (MCP) specification, without exposing external network ports. To ensure control over the process, a dynamic interruption system (Approval Gating) has been implemented, supporting three levels of autonomy (Manual, Semi-auto, Full-auto) and allowing the specialist to approve high-risk actions. To guarantee the immutability of the evidence base, cryptographic signing of the obtained artifacts using the Ed25519 algorithm is implemented. The conclusions prove that the proposed architecture eliminates the problem of excessive artificial intelligence agency, creating a reliable intermediary layer between the target system and the expert. Future research will focus on scaling the infrastructure to enable parallel testing of a significant number of hosts.