With the advancement of Large Language Models (LLMs), increasingly
sophisticated and powerful GPTs are entering the market. Despite their
popularity, the LLM ecosystem still remains unexplored. Additionally, LLMs'
susceptibility to attacks raises concerns over safety and plagiarism. Thus, in
this work, we conduct a pioneering exploration of GPT stores, aiming to study
vulnerabilities and plagiarism within GPT applications. To begin with, we
conduct, to our knowledge, the first large-scale monitoring and analysis of two
stores, an unofficial GPTStore.AI, and an official OpenAI GPT Store. Then, we
propose a TriLevel GPT Reversing (T-GR) strategy for extracting GPT internals.
To complete these two tasks efficiently, we develop two automated tools: one
for web scraping and another designed for programmatically interacting with
GPTs. Our findings reveal a significant enthusiasm among users and developers
for GPT interaction and creation, as evidenced by the rapid increase in GPTs
and their creators. However, we also uncover a widespread failure to protect
GPT internals, with nearly 90% of system prompts easily accessible, leading to
considerable plagiarism and duplication among GPTs.

该研究论文探索了 GPT 商店的漏洞和抄袭问题，发现了大规模监控和分析商店以及提取 GPT 内部的有效工具，并揭示了 GPT 内部保护失效导致的广泛抄袭问题。