Taxonomy sources.xlsx

We conducted a thorough literature review with the goal to develop a taxonomy of attacks towards ChatGPT. The main search engine used to search for relevant literature was Google Scholar. Our search included the keywords ChatGPT or LLM combined with Attack, Jailbreak or any of our attack types mentioned in the taxonomy. For our taxonomy we ended up with 15 relevant studies all mentioning either one of or multiple of our topics: Prompt Manipulation, Output Misappropriation, Information Extraction and Privacy Breaches.

本研究开展了全面的文献综述，旨在构建针对ChatGPT的攻击分类体系。本次综述主要依托谷歌学术（Google Scholar）作为检索相关文献的核心搜索引擎。本次检索采用的关键词组合为：以ChatGPT或大语言模型（LLM）搭配“攻击（Attack）”、“越狱（Jailbreak）”，或本分类体系中提及的任意攻击类型。最终，本分类体系共纳入15篇相关文献，这些文献均涉及本研究定义的至少一个或多个主题：提示词操纵（Prompt Manipulation）、输出盗用（Output Misappropriation）、信息抽取（Information Extraction）以及隐私泄露（Privacy Breaches）。