The Poorest Man in Babylon: A Longitudinal Study of Cryptocurrency Investment Scams

Muhammad Muzammil, Abisheka Pitumpe, Xigao Li, Amir Rahmati, Nick NikiforakisWWW 2025 · OpenReview

TL;DR

Researchers have developed Crimson, the first large-scale system to detect cryptocurrency investment scam websites in real-time by monitoring the "birth" of websites via Certificate Transparency logs. Over an eight-month period, the system identified more than 43,000 unique scam sites, revealing a massive, centralized infrastructure where a tiny fraction of hosting providers and IP addresses support the majority of these fraudulent operations.

What Problem Does This Paper Address?

Cryptocurrency investment scams are a multi-billion dollar problem, yet they are notoriously difficult to track at scale. Unlike "giveaway scams" (which simply post a wallet address on a homepage), investment scams use professional-looking websites that require users to create accounts and log in before revealing deposit addresses.

Previous research often relied on finding scam links manually on social media or YouTube comments. This creates a "bottleneck of discovery." This paper addresses the need for a proactive, automated way to find these sites the moment they are created, understand their infrastructure, and estimate the actual financial damage they cause.

Methods

The authors built a distributed pipeline called Crimson that consists of several sophisticated stages:

  1. Domain Selection: It monitors Certificate Transparency (CT) logs—public records of every TLS certificate issued. By processing ~1 million certificates per hour and using keyword segmentation (splitting btcethinvestments.com into btc, eth, investments), it flags potential scam domains.
  2. Content-based Filtering: For flagged domains, Crimson waits 12 hours (to allow the site to go live), takes a screenshot, performs OCR (Optical Character Recognition) to extract text, and checks for "scammy" keyword clusters (e.g., "high returns," "secure," "withdraw").
  3. Hybrid LLM Classification: To avoid manual labeling, the system uses Large Language Models. It primarily uses Llama 3:70b for cost-effectiveness and routes "inconclusive" cases to GPT-4 for a final verdict.
  4. Automated Account Creation: This is the most novel technical hurdle. Crimson simulates a user by signing up for accounts and navigating internal dashboards to find the hidden cryptocurrency wallet addresses where victims are told to send money.
  5. Clustering & Analysis: The system uses perceptual hashing (p-hash) to group sites with identical visual designs and analyzes blockchain transactions to track the flow of stolen funds.

Datasets

  • CT Logs: The primary source, processing approximately 6 billion domain names over 8 months.
  • URLScan: Used to compile an initial training set of known scam keywords.
  • Tranco Top 10K: Used as a "benign" baseline to compare top-level domain (TLD) and keyword usage.
  • Crimson Dataset: A newly created dataset of 43,572 unique scam websites and their associated metadata (IPs, wallets, HTML), which the authors plan to open-source.

Key Insights and Ideas

  • Infrastructure Concentration: 52% of all scam websites were hosted on just 10% of the resolved IP addresses. This suggests that a small number of "bulletproof" or low-cost hosting providers (like Hostinger) are the backbone of the scam industry.
  • The "Lazy Scammer" Effect: 88% of detected sites shared commonalities in design, JavaScript, or contact info. Scammers are heavily reusing templates, making them vulnerable to cluster-based blocking.
  • Blacklist Failure: Popular security tools (Google Safe Browsing, MetaMask, etc.) are failing. The vast majority of Crimson-detected sites were absent from these blacklists, meaning users visiting these sites receive no warning.
  • Shared Wallets: In a "smoke and mirrors" revelation, the researchers found that many sites show the same wallet address to different users. This proves the sites aren't even bothering to track individual "investments"—they are simply collection buckets.

Key Results

  • Scale: Crimson detected 43,572 unique scam websites between January and September 2024.
  • Accuracy: The GPT-4 classification model achieved 90% accuracy in identifying scams.
  • Financial Impact: By analyzing just 6.7% of the detected sites (where wallet extraction was successful), the researchers confirmed $2.04 million USD** in losses. Extrapolating this to the full dataset suggests total losses well **exceeding $100 million USD.
  • Persistence: 47% of the detected scam sites remained active through the end of the study, showing that these operations are remarkably resilient to current takedown efforts.

Limitations and Future Work

The authors acknowledge several challenges:

  • Evasion: If scammers stop using specific keywords (like "crypto" or "invest") in their domain names or homepage text, they could bypass Crimson’s initial filters.
  • Anti-Bot Measures: Many scam sites use CAPTCHAs or manual admin approval for new accounts, which hinders Crimson’s ability to automate wallet extraction.
  • LLM Costs: While Llama 3 is free to run locally, the hybrid approach still incurs costs when GPT-4 is required for high-confidence verification.

Future work could involve expanding the system to detect scams on alternative blockchains (beyond Bitcoin and Ethereum) and developing more robust ways to bypass anti-bot protections during the account creation phase.

Summary updated

Abstract

Governments and regulatory bodies have recognized investment scams as the most prevalent forms of cryptocurrency fraud. These scams typically use professional-looking websites to lure unsuspecting victims with promises of unrealistically high returns. In this paper, we introduce Crimson, a distributed system designed to continuously detect cryptocurrency investment scam websites as they are created in the wild. Over the first 8 months of 2024, Crimson processed approximately 6 billion domain names and classified 43, 572 unique cryptocurrency investment scam websites in real-time. Beyond detection, we provide insights into the design and infrastructure of these websites that can help users recognize scam patterns and assist hosting providers in detecting and blocking such sites. Among others, we discovered that most investment scam websites use similar templates and that 52% of all scam websites were hosted on just 10% of all resolved IP addresses, indicating a concentration of scam operations within a small subset of hosting providers. Furthermore, we investigate the inclusion of our detected scam websites in blacklists used by popular web browsers and applications, finding that the vast majority of these websites were absent. On the financial side, by analyzing the incoming transactions to scammer wallets on 6.7% of the sites detected by Crimson, we observe an estimated lower bound of 2.04M USD in losses because of cryptocurrency investment scams, pointing to tens of millions of dollars of losses in total.
BibTeX
@inproceedings{muzammil2025poorest,
  author = {Muhammad Muzammil and Abisheka Pitumpe and Xigao Li and Amir Rahmati and Nick Nikiforakis},
  title = {{The Poorest Man in Babylon: A Longitudinal Study of Cryptocurrency Investment Scams}},
  booktitle = {The Web Conference},
  year = {2025},
  url = {https://openreview.net/forum?id=P0x8J5gCPP},
}
Figure