Cybersecurity-Ready Agentic AI for SOC Automation: A Safe Multi-Agent Framework with Human-in-the-Loop Escalation
Abstract
The paper is an agentic artificial intelligence architecture of Security Operations Center (SOC) automation that meets the challenges of signature based intrusion detection to large scale, multi-vector cyber threats. Current SOC systems have deficiencies in alert fatigue, absence of contextual reasoning, and uncontrolled automation limits with clear safety concerns. It provides a four-agent architecture that includes the Log Analysis agent, Threat Correlation agent, Threat Enrichment agent and Threat Escalation agent, which are the end-to-end governed pipeline that is not discussed by the previous work. The framework uses an XGBoost and Synthetic Minority Oversampling Technique as the main detector, compared to Random Forest and Isolation Forest, and presents a bounded-autonomy escalation policy that uses confidence and novelty thresholds to apply uncertain or high-severity cases to human analysts. The results of experiments with the CIC-IDS-2017 dataset (500,000 records) indicate that XGBoost has an accuracy of 98.87, F1-score of 0.9617, and inference latency of 318.688ms. The agentic pipeline results in 17,348 correlated incidents, which directs 4,265 autonomously and escalates 13,083 incidents to human inspection at 93.81% incident accuracy. Robustness analysis demonstrates that sensitivity to perturbation of features is very high, highlighting the necessity of adversarial hardening. Flow Duration, Flow Packets per second and Total Forward Packets are found to be the most significant drivers of escalation identified as SHapley Additive exPlanations-based interpretability. Results confirm that controlled-autonomy SOC operation leads to high detection performance in a way that maintains human oversight and operational safety.