Adaptive Cognitive Profiling for Executive AI Agents Amid Emerging AI Impersonation Threats

Abstract

Rapid digitalization and the adoption of large language models in executive workflows have heightened cybersecurity risks, particularly AI-enabled impersonation and authority abuse. This study develops and evaluates an adaptive cognitive profiling framework for executive AI agents that integrates behavioral biometrics, anomaly detection, and continuous trust assessment to enhance identity assurance and governance. The framework combines keystroke-dynamics modeling using Mahalanobis distance classification, Isolation Forest–based anomaly detection, and a continuous adaptive trust score fusing behavioral, cognitive-decision, contextual, and anomaly signals. Evaluation was conducted using a synthetically generated keystroke dataset modeled on the CMU benchmark, comprising 8,000 samples from 20 subjects across multiple sessions, with experiments performed in a controlled, simulated environment. Results demonstrate an authentication accuracy of 93.78%, an equal error rate of 5.99%, anomaly detection recall of 95.52%, and successful identification of 93.75% of simulated impersonation attacks. Governance mechanisms incorporating human-in-the-loop authorization and graduated access control operationalize principles from the NIST AI Risk Management Framework. Key limitations include reliance on synthetic behavioral data, elevated false-positive rates in anomaly detection, and the absence of longitudinal real-world executive validation. Future work will focus on multimodal data integration, long-term behavioral drift analysis, and deployment within live enterprise environments.