The Evolving Landscape of Cybersecurity: Red Teams, Large Language Models, and the Emergence of New AI Attack Surfaces

Abstract

This study explores cybersecurity questions using a question-and-answer format with the advanced ChatGPT model from OpenAI. Unlike previous chatbots, ChatGPT demonstrates an enhanced understanding of complex coding questions. We present thirteen coding tasks aligned with various stages of the MITRE ATT&CK framework, covering areas such as credential access and defense evasion. The experimental prompts generate keyloggers, logic bombs, obfuscated worms, and ransomware with payment fulfillment, showcasing an impressive range of functionality, including self-replication, self-modification, and evasion. Despite being a language-only model, a notable feature of ChatGPT showcases its coding approaches to produce images with obfuscated or embedded executable programming steps or links.