Yuval Elovici

Threat modeling is a cornerstone of proactive cloud security, yet traditional frameworks such as STRIDE rely on manual analysis that cannot scale to meet the demands of modern infrastructures. Recent approaches leverage large language models (LLMs) to automate threat identification, but single-model frameworks exhibit inconsistent performance and lack contextual enrichment. To overcome these limitations, we introduce ATMoCloud, a multi-agent framework that automates end-to-end threat modeling for cloud-based applications. ATMoCloud incorporates several design elements customized for security analysis, orchestrating specialized agents that perform the following in its service pipeline: (1) preprocessing, (2) dual-stage enrichment (component and threat level), (3) validation, (4) STRIDE-based threat generation, (5) quantitative risk scoring, and (6) actionable mitigation synthesis. By distributing these tasks across cooperative agents, the framework prevents context exhaustion and maintains reasoning consistency. We evaluated ATMoCloud on 50 real-world architectures from three major cloud providers (AWS, Azure, GCP), performing semantic coverage analysis, a five-judge comparative assessment, ablation studies, failure analysis, and human-expert validation. ATMoCloud achieved strong alignment with expert models (0.78 similarity) while generating validated novel threats (93.33% relevance), outperforming both automated and human-curated baselines. By completing analyses in 20 minutes, compared to dozens of hours spent on manual analysis, ATMoCloud demonstrates that multi-agent collaboration enables scalable, high-precision threat modeling, providing actionable security insights.