Author ORCID Identifier

https://orcid.org/0000-0002-3455-2562

Document Type

Article

Publication Date

Fall 11-4-2026

Department

Department of Computer Science

Abstract

Adaptive AI systems are changing adversarial interaction because defensive actions no longer merely affect an environment; they also become observable signals that adversaries can interpret, remember, and exploit. In cybersecurity settings such as phishing mitigation, fraud detection, and attack surface management, a defender’s probes, friction, and redirection mechanisms may therefore reveal as much as they protect. This paper argues that detectability must be treated as a first-order design constraint in AI-mediated defense rather than an implementation detail or secondary usability cost. Cybersecurity is our primary domain, but the constraint is general to AI-mediated adversarial interaction. Any defensive intervention exposed to an adaptive agent can become part of what that agent learns. We introduce a taxonomy distinguishing two core interaction-based defenses: Covert Interaction Testing (CIT), which subtly probes adversary characteristics such as memory and adaptivity, and Adversarial Resource Shaping (ARS), which strategically imposes operational costs to degrade attacker efficiency. We formalize the tension between defensive utility and exposure risk as the stealth-effectiveness frontier, bounded additionally by the unavoidable friction imposed on benign users. Finally, we present a minimal operational model of repeated interaction to derive open research problems, calling for dynamic, fast-adapting defense mechanisms capable of operating under continuous adversarial observation.

Share

COinS