Introduction

This repo contains the code and information for the paper "Anatomy of an AI-powered malicious social botnet".

We provide a case study on a Twitter botnet, i.e., the fox8 botnet, that uses ChatGPT to generate negative/harmful content and promote suspcious websites.

Data release

We release the fox8-23 benchmark dataset for distinguishing LLM-powered social bots and humans. The raw data can be downloaded from zenodo.

The dataset contains the following accounts:

Account type	Source	Number
Bot	fox8 bots (this work)	1140
Human	botometer-feedback	285
Human	gilani-17	285
Human	midterm-2018	285
Human	varol-icwsm	285

For each account, we share up to 200 tweets from it. The tweet objects are in Twitter API V1.1 format. Each line of the raw data file is a JSON object with the following schema:

{
  "user_id": 123456,
  "label": "bot",
  "dataset": "fox8",
  "user_tweets": [
    tweet1,
    tweet2
    ...
  ]
}

Code

We also share the code used to query OpenAI's AI text editor in query_openai_detector.py.

Citation

@article{yang2024anatomy,
  title={Anatomy of an AI-powered malicious social botnet},
  author={Yang, Kaicheng and Menczer, Filippo},
  journal={Journal of Quantitative Description: Digital Media },
  volume={4},
  url={https://journalqd.org/article/view/5848},
  DOI={10.51685/jqd.2024.icwsm.7},
  year={2024},
  month={may}
}