Outlet Title

Issues in Information Systems

Document Type

Article

Publication Date

2026

Abstract

Phishing remains a leading initial-access vector despite sustained investment in user training and technical filtering. The literature on training decay, bounded technical controls, and adversarial adaptation each treat one slice of the problem in isolation. We develop a discrete-time simulation that couples all three subsystems within a unified socio-technical framing, deliberately modeling the adversary as a minimally adaptive multi-armed bandit. The choice is methodological: a minimal-adaptation adversary establishes a conservative lower bound on adversarial capability, so that more sophisticated attackers, including those driven by large language models, can be expected only to elevate the resulting steady-state loss floor. Across 200 random seeds and a 21-point control-strength sweep, the floor is structurally positive in every condition tested; the adaptive premium over a uniform-random baseline is approximately 12% and persists across alternative learning algorithms, feedback structures, and an enlarged action space; and the loss- versus-control-strength relationship is approximately linear in residual exposure. The bandit’s selections concentrate on high-quality, high-personalization actions, the operational mechanism through which AI- enabled adversaries elevate the floor in our framework. We interpret the floor as the planning baseline against which the agentic-AI adversarial landscape should be evaluated, not as a ceiling.

Download

Share

COinS