Keerthana Madhavan (School of Computer Science, University of Guelph, Canada), Luiza Antonie (School of Computer Science; CARE-AI, University of Guelph, Canada), Stacey D. Scott, School of Computer Science; CARE-AI, University of Guelph, Canada)
Election security Security Operations Centers (SOCs) face an expanding mandate: beyond traditional network defense, they must now detect cognitive threats, content that manipulates audiences through psychological tactics rather than explicit falsehoods. Existing tools provide binary labels without explaining how manipulation occurs, limiting triage and response. We present E-MANTRA, a Large Language Model (LLM)-based framework that integrates agentic Artificial Intelligence (AI) into SOC workflows by identifying six manipulation tactics (emotional manipulation, conspiracy framing, discrediting, trolling, impersonation, polarization) with explainable classifications. Evaluated on 900 election-related samples, E-MANTRA attains 54.2% triage accuracy and an estimated 57% workload reduction under confidence-based decision-making. Results confirm exploitable model specialization: Llama-3 70B excels at conspiracy detection (F1=0.71), GPT-3.5 at emotional manipulation (F1=0.66), Mistral- Small at discrediting (F1=0.63). Category-aware routing improves accuracy by 2.4 percentage points over the best single model at $0.005 per classification. We provide a practitioner-oriented deployment checklist, cost models, and Security Information and Event Management (SIEM)/Security Orchestration, Automation, and Response (SOAR) integration guidelines to support operational adoption in election security SOCs.