US23-Heiding-Devicing-and-Detecting-Phishing.pdf

This Black Hat USA 2023 presentation analyzes the risks and capabilities of using large language models (LLMs) to craft and automate phishing attacks. The researchers compare LLM-generated emails, a human-centered V-Triad model, and combinations thereof in an experiment with Harvard-affiliated participants, reporting click-through rates as high as ~65% for V-Triad messages and elevated success when using targeted, credibility-focused templates. The talk outlines a four-phase automation workflow (collect background information, create phishing emails, send emails, analyze results), gives example phishing templates, evaluates intent/AI-detection by various models, and concludes with both warnings about easier-to-launch sophisticated phishing and recommendations to leverage LLMs defensively for training and detection.