Phishing Detection Robustness Collapses Under Adversarial Distribution Shift: Attack-Surface Concentration

arXiv 2603.19204·medium signal

This paper (arXiv:2603.19204) shows that phishing detectors built on engineered website features achieve near-perfect i.i.d. accuracy but have concentrated attack surfaces that adversaries can cheaply exploit. Deployment security depends critically on how attackers adapt, and the paper demonstrates that robustness and cost are in tension with feature-based approaches. Relevant for security practitioners embedding ML-based detection in LLM agent pipelines.