Evaluating Large Language Models’ Ability to Automate Spear Phishing
Expert Systems with Applications, v. 314, June 5, 2026, 131546.
EXCERPT:
In this paper, we investigate the dual-use nature of large language models (LLMs) in the phishing domain, evaluating both their offensive and defensive capabilities. We first assess LLMs’ capacity to automate personalized spear phishing attacks, comparing their performance with human experts across N = 101 participants in four experimental groups: control (12% click-through), human experts (54%), fully AI-automated (54%), and AI with human-in-the-loop (56%). The automated tool produced accurate target profiles in 88% of cases. We then evaluate LLMs’ defensive potential for phishing detection, testing Claude 3.5 Sonnet across 381 emails and achieving 97.25% detection accuracy with zero false positives. Economic analysis reveals that AI automation increases phishing profitability by up to 50× for large-scale campaigns. These findings highlight both the threat posed by AI-automated phishing and the promise of AI-powered defenses, underscoring the need for balanced offensive-defensive strategies in an AI-enabled threat landscape.
Categories: AI/LLMs