Evaluating large language models’ ability to automate spear phishing

doi:10.1016/j.eswa.2026.131546

Evaluating large language models’ ability to automate spear phishing

2026-02-06

0

PRE

AI

Claudio Mayrink Verdun

DOI：10.1016/j.eswa.2026.131546

Original

Original request for help

Save

Abstract

En 中文

• Examines fully automated AI spear-phishing attacks validated on 101 human subjects. • AI phishing emails achieved  ∼ 54–56% click-through, on par with human experts. • Economic analysis shows AI phishing automation can boost attacker ROI by up to 50 × . • Human-subject study approved by university IRB and ethically reviewed. • Priming models for suspicion improved detection accuracy while maintaining low FP.

Keywords:

AI spear-phishing

large language models

click-through rate

attacker ROI

human subject study

Evaluating large language models’ ability to automate spear phishing

Abstract

Journal

Organization