Multi-Agent System for Automated Red Teaming Workflows: Methodology and Architecture
Year: 2026
Type: conference-paper
Abstract: Manual penetration testing can no longer keep pace with modern attack surfaces and AI-accelerated adversarial threats. To this end, we propose an automated red teaming methodology structured across three phases: Reconnaissance, Execution, and Post-Execution in alignment with the National Institute of Standards and Technology Cybersecurity Framework 2.0 and the European Cybersecurity Skills Framework to enable adoption within existing security organizations. Our methodology is implemented via a hierarchical multi-agent architecture, where a supervisor agent coordinates specialized sub-agents under bounded execution contexts. Evaluation occurs in a dockerized environment using phase-specific verifiable scoring functions, with experiments utilizing both proprietary and open-source LLMs, benchmarked against a single-agent baseline. Results show consistent performance gains with the multi-agent architecture over the single-agent architecture with the execution phase representing a bottleneck for open-source models. The work establishes a foundation for scalable, standards-aligned deployment of automated red teaming workflows in operational cybersecurity settings.
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225624403