Devising and Detecting Phishing: Large Language Models vs. Smaller Human Models
summary
A controlled study sends tailored emails to 112 participants and compares GPT‑4 generated, human‑crafted V‑Triad, hybrid V‑Triad+GPT, and generic control messages. Results show click‑through rates of control 19–28%, GPT 30–44%, V‑Triad 69–79%, and V‑Triad+GPT 43–81%, demonstrating that small, psychology‑driven templates outperform vanilla LLM output but that semi‑automation reaches similar effectiveness; participants explain their clicks, revealing mixed motivations and enabling LLM‑based detection and training tools.
AI-written summary · about
chapters (14)
- 0:00AI-driven phishing threats
- 3:15AI for good and deception
- 5:43Large vs. human models
- 8:08GPT simplifies phishing
- 11:13Collecting phishing data
- 13:20Automated vs. manual phishing
- 16:10Phishing email study results
- 18:16Participant feedback
- 21:05Detecting phishing intentions
- 24:54AI-generated phishing emails
- 27:21AI browser extensions for defense
- 28:26Future AI cybersecurity research
- 32:15Call to action
- 32:57Key takeaways
related talks
34mCan a General-Purpose LLM Do Tier-1 Triage?
46mHow Large Language Models Are Reshaping the Cybersecurity Landscape
49mChatGPT: A New Generation of Dynamic Machine Based Attacks?
42mGenerative Email Attacks and How to Defend Against Them
51mLarge Language Models: Friend, Foe, or Otherwise
52mUnleash the Power of Large Language Models (LLMs)
27mBlunting the Phisher's Spear: A Risk-Based Approach...
48mGPT3+ Models in Cybersecurity Tasks