- The paper introduces Anansi, a semi-automated pipeline combining persistent LLM personas, browser automation, and cryptocurrency tracing to engage 1,901 scammers across 29,209 messages from 7,028 phone numbers.
- The study identifies extensive reuse across scam templates, websites, and infrastructure, while documenting domain fronting and rapidly rotating wallet addresses that complicate campaign attribution and takedowns.
- The paper traces more than $12 million in estimated cryptocurrency losses and finds that mainstream blocklists detected only 0–29.6% of 135 scam domains, exposing major gaps in current defenses.
Overview
The paper presents [SYSTEM], a semi-automated measurement pipeline for characterizing job-based smishing scams—a subset of "pig butchering" fraud in which victims are recruited under the pretense of paid remote micro-tasks and progressively induced to deposit cryptocurrency. Over a 10-month measurement period (April 14, 2025 – January 31, 2026), the system collected 29,209 messages from 7,028 scammer phone numbers across SMS, WhatsApp, and Telegram, engaged 1,901 scammers in conversation, and traced funds flowing to 354 scammer-controlled Bitcoin and Ethereum wallets. The authors report an estimated **$12M+ in cryptocurrency losses** attributable to the wallet addresses they extracted, with individual Ethereum operations ranging up to roughly$2 million per wallet. The work's principal contribution is methodological: it demonstrates that LLM-driven persona management combined with automated browsing can sustain direct engagement with scammers at a scale previously unattainable through manual interaction studies.
Motivation and threat model
Job scams follow a staged social engineering sequence: unsolicited contact promising $250–$500 daily for minimal remote work; trust building via HR personas; handoff to a "trainer" on WhatsApp or Telegram; registration on a fraudulent task platform; completion of review or engagement tasks that display fictitious earnings; and finally payment extraction via minimum-balance and "prepay" requirements before the scammer ghosts the victim. The economic context is substantial: the authors cite FBI figures of approximately $2.8 billion lost to cryptocurrency fraud in the US in 2024 (a 71% year-over-year increase), over$5.8 billion in pig-butchering losses per Operation Level Up, and FTC-reported job-scam crypto losses of $41 million in the first half of 2024 alone—already double all of 2023.
Prior measurement work has either been retrospective (SMS phishing corpora [(Clerc et al., 2024)-style analyses are cited as Nahapetyan et al. and Agarwal et al.]), web-centric (cryptocurrency investment scam detection), or manual and small-scale (telephone and YouTube scammer interactions). None offered a real-time pipeline combining message capture, active engagement, multimodal website analysis, and on-chain financial tracing.
System design
[SYSTEM] comprises five modules. Data acquisition scrapes public reporting portals—the BBB Scam Tracker and Smishtank—via Selenium, supplemented by OCR-processed screenshots submitted by collaborators; 6,874 numbers came from public portals and 154 from direct submissions. LLM-assisted engagement initiates contact via Twilio programmable SMS/WhatsApp and the Telegram bot API, assigning each conversation a persistent synthetic victim persona so that demographic details remain consistent even when conversations migrate across platforms or between scammer handlers. Human oversight is triggered when generated replies risk appearing mechanical. Task completion uses Selenium and ChromeDriver to register accounts on scam platforms, complete assigned tasks, and capture screenshots. Wallet extraction contacts each scam site's customer-support channel—using OpenAI's Operator plus custom scripts—to explicitly request deposit addresses. An analysis layer then clusters messages, infrastructure indicators, and financial flows.
A notable operational constraint is Meta's WhatsApp Business API policy forbidding business-initiated conversations with users who have not messaged first. The authors' workaround—asking the original scammer contact to relay trainer details or migrate to Telegram—shifted observed handoffs toward Telegram (294 conversations) relative to WhatsApp (53), which likely biases platform-trajectory measurements away from WhatsApp's true popularity among scammers.
Engagement funnel and message characteristics
Of 7,028 contacted numbers, 1,944 failed delivery (693 unreachable handsets, 692 landlines, 141 WhatsApp restrictions); 1,901 scammers responded. The response-to-engagement attrition implies that roughly 27% of contacted numbers yielded substantive interactions, which the authors treat as evidence that mass-messaging scammers respond indiscriminately to any reply—they engage with responses to messages originally sent to other victims, indicating no recipient tracking.
Initial outreach falls into three categories: detailed job offers, exploratory recruiter inquiries, and wrong-number texts. Wrong-number engagements proved largely unsuccessful to recruit into, a limitation of the engagement methodology for that vector. After preprocessing out stopwords, person names, and brand names, clustering revealed heavy template reuse: the largest cluster contained 600 identical initial messages. Brand impersonation is pervasive—Target (454 mentions), Amazon (417), Costco (385), LinkedIn US (357)—and personas skew female, led by the recurring alias "Jasmine Martine" (312 occurrences). Of 25 Telegram profile pictures tested against an AI-image detector, only 3 were flagged as AI-generated; the remainder traced to legitimate social media profiles, contradicting any assumption that generative imagery dominates scammer persona construction.
Persistence analysis shows 89% of scammers cease messaging within one day of last contact and ~98% within ten days, but a persistent tail (~0.5%) continues outreach beyond a month.
Infrastructure analysis
Analysis of 135 unique scam websites reveals pronounced infrastructure concentration: Cloudflare-hosted networks account for 44.8% of sites, and twelve sites share a single INTEGEN-2 IP address (192.252.179.27) in Japan. TLD distribution diverges sharply from the Tranco Top 1M: .vip appears on 16.67% of scam sites versus 0.31% of Tranco domains, and .lat on 6.86% versus 0.03%—consistent with prior findings on low-cost, laxly regulated registries.
The most operationally interesting finding is JavaScript-based domain fronting. A domain observed under maintenance redirects each reload to a distinct operational subdomain clustered under two second-level domains sharing an IP, all landing on the same login path (/index/user/login.html). This defeats static URL analysis, complicates takedowns, and means IOC-based clustering almost certainly underestimates shared infrastructure.
Clustering across three modalities corroborates coordination: template reuse (e.g., a training-code message appearing verbatim in 113 conversations), web design reuse (a single homepage template across 42 of 218 sites in the largest cluster), and indicator overlap (one Structube-impersonating subdomain used by 13 scammers). Cryptocurrency address clustering, however, is undermined by deliberate rotation: in a controlled experiment, a scammer supplied a fresh address per request and stated that "each cryptocurrency transaction address has a 30-minute validity period." This is a direct challenge to prior work treating wallet co-occurrence as a reliable campaign-linkage signal.
Task types split into product-review manipulation (67.5%), YouTube engagement (18.3%), and app-store manipulation (14.2%). Small payouts ($3–$15) flow through regulated services—Cash App (286 unique conversations), PayPal (183), Coinbase (54)—implying stolen identities, forged KYC documents, or money mules, though the paper does not determine which.
Financial loss estimation
Following Gomez et al.'s CCS 2023 methodology, the authors trace inbound transactions via Blockstream and Etherscan APIs, exclude intra-dataset transfers, filter same-day refund-like outbound flows, and convert to USD at historical rates. From 262 Bitcoin and 209 Ethereum addresses collected (205 and 149 unique after reuse), Bitcoin wallets showed a median of 11 transactions each, while Ethereum wallets showed far higher activity—a median of 34 and a maximum of 5,090 transactions on a single wallet. The single highest-revenue Ethereum operation (pzmsa-751.com / tx11.vip) received approximately **$2.075 million across 5,090 transactions**, and several other wallets exceed$800K. These figures represent lower bounds insofar as they cover only wallets the pipeline successfully extracted, but they also inherit uncertainty from exchange-rate timing and from the heuristic filtering of refunds—an inherent limitation of blockchain-attribution approaches the authors acknowledge implicitly through their methodology choices rather than quantify.
Blocklist coverage
Cross-checking the 135 scam domains against commercial blocklists yields stark results: only 29.6% appear in VirusTotal, 15.6% in Google Safe Browsing, 7.78% in MetaMask's blocklist, and 0% in PhishFort. This gap is structural rather than incidental—job scam sites contain no malicious code by design, since their legitimacy is the attack surface—and indicates that current detection infrastructure optimized for malware and phishing systematically misses this category.
Limitations
The authors are explicit about three constraints. First, the pipeline remains semi-automated: transitions into task completion and customer-support contact require manual configuration, and LLM conversations receive human monitoring; scalability derives partly from natural attrition narrowing thousands of contacts to a manageable subset. Second, WhatsApp API restrictions bias both reachability and observed platform preferences. Third, coverage is limited to task-type job scams within the United States, excluding romance, delivery, investment-only, and tech-support variants, as well as cross-border dynamics where operations span jurisdictions with different languages and cultural targeting.
An additional caveat not fully quantified: because the system responds to messages reported by third parties rather than messages sent to its own numbers, the sample conditions on victim reporting behavior, and public portal latency introduces a temporal lag between scammer activity and observation.
Conclusion
This paper establishes that end-to-end automation of scammer engagement—LLM personas, scripted browser agents, and support-channel probing—can produce longitudinal behavioral, infrastructural, and financial data on job scams that passive corpus analysis cannot. Its strongest empirical claims are the scale of template and infrastructure reuse, the demonstrated short-horizon wallet rotation that undermines address-based clustering, and the near-total absence of these scams from mainstream blocklists. The open questions it leaves include whether fully autonomous engagement can match human-supervised success rates, how to link campaigns behind domain-fronted infrastructure, and how the ecosystem's economics differ outside the US reporting ecosystem from which its sample is drawn.