Duo Security researchers have identified a large-scale botnet on Twitter dedicated to cryptocurrency scams. The discovery was part of a broader effort to develop methods for detecting automated accounts on the platform.

Using Twitter’s API, the team compiled a dataset of approximately 88 million public accounts and more than 500 million tweets. They focused their analysis on the last 200 tweets from each account. Applying traditional machine learning techniques, they trained a classifier to distinguish between human and bot accounts. Additional data science approaches helped map and analyze the botnet structures.
Their goal extends beyond detection; they aim to promote further research into automation on Twitter. By open-sourcing their methodology and data collection process, they seek to encourage other investigations into differentiating genuine accounts from automated or hybrid accounts, which combine automation with human input.
The study primarily targeted ‘pure’ bots rather than hybrid accounts, which are more complex and often involve human oversight. While sentiment analysis was not the focus, the researchers acknowledged that hybrid accounts, such as customer service bots, may have been included in their dataset. These often pose greater challenges for detection and are sometimes used in disinformation campaigns.
A key case study highlighted a crypto scam botnet estimated to comprise at least 15,000 bots, though the real number could be significantly higher. These bots promote schemes that direct users to malicious links, promising giveaways or other incentives to steal money. Despite notifying Twitter about the botnet, the researchers found it remains active, impersonating legitimate accounts—including those of news agencies—and still recommended followings of other spam bots through Twitter’s algorithms.
When asked about efforts to combat such automation, a Twitter spokesperson stated that the platform employs various detection techniques and has measures to address deceptive practices. They claimed that less than 5% of accounts are involved in spam activities and added that while some automated accounts are benign, others serve malicious purposes.
The spokesperson referenced positive uses of automation, citing examples like Pentametron, a bot that creates rhyming tweets in iambic pentameter, illustrating the nuanced debate over bots' roles on the platform.
The research highlights the importance of accurately discriminating between genuine users and automated accounts, especially as disinformation tactics evolve. The team suggested that future platform features, such as a ‘bot score,’ could help users identify potential scams more readily—though no official plans for such a feature were announced.
Challenges faced included limited access to data supporting existing detection methods. Data scientist Olabode Anise noted frustrations with transparency issues and the tendency for prior studies to jump to conclusions without foundational data. Their efforts aim to create a robust base for subsequent research, prioritizing transparency and thoroughness.
This work underscores the complexity of social media automation and the ongoing need for sophisticated detection tools. As the landscape of online misinformation and scams expands, these findings provide a critical step toward more effective strategies to protect users and preserve the integrity of online interactions.
Corrections and clarifications: contact@avorz.com
How this was reported: /how-we-report/