ISoLA & AISoLA 2026

International Symposium on Leveraging Applications of Formal Methods, Verification, and Validation • Kos, Greece

Talk

Policy-as-Code Authorization for Multi-agentic Pipelines in Security Operations Centers

Time: Friday, 30.10

Room: Room 2

Authors: Ekaterini Gougousi, Anna Gazani, Spyridon Kounoupidis, Miltiadis Siavvas, Panagiotis Katsaros

Abstract: Security Operations Centres (SOCs) are critical for defending organizations against increasingly complex cyber attacks. However, the overwhelming volume of security events often floods these environments, creating high-stress conditions that delay incident triage, analysis, and mitigation. To effectively address this, SOCs must be recognized as complex sociotechnical systems where operational success relies not only on deployed technologies but fundamentally on the cognitive capacity, expertise, and decision-making performance of human analysts. While Generative Artificial Intelligence (GenAI) and Large Language Model (LLM) based agents offer a powerful mechanism to reduce analyst workload and decrease the Mean Time to Response (MTTR) by automating routine tasks, their deployment must be strictly governed by this human-centric framework. In this paper, we present a novel framework that integrates multi-agentic GenAI pipelines into the SOC environment while preserving the human analyst as the ultimate decision-maker and validator. Our primary contribution is a policy-as-code authorization scheme that ensures autonomous AI agents—acting on behalf of Level 1 (L1), Level 2 (L2), or Level 3 (L3) analysts—strictly adhere to role-based permissions. Leveraging the Open Policy Agent (OPA), the proposed mechanism evaluates an agent's intended action plan against machine-readable policies prior to execution. If an agent attempts an action exceeding its authorized privileges, the execution is blocked, and an automatic escalation to a higher-tier analyst is triggered. Furthermore, to guarantee trusted AI utilization, we introduce a mandatory Human-In-The-Loop (HITL) mechanism, ensuring that critical mitigation actions (e.g., endpoint quarantine, subnet isolation) are never executed without explicit human approval. We evaluated our approach in a real-world SOC setting with actual L1 and L2 analysts responding to active incidents. The results demonstrate a significant reduction in MTTR while successfully preventing unauthorized AI executions. Ultimately, this work provides a secure pathway for adopting GenAI in SOC pipelines, utilizing technology to alleviate tedious workloads while reinforcing the core sociotechnical paradigm: AI must support, rather than replace, human expertise, coordination, and judgment.