The Imitation Game: Using Large Language Models as Chatbots to Combat Chat-Based Cybercrimes

1 INTRODUCTION

• Traditional filters struggle with context-dependent, multi-message scams and multimedia such as screenshots or QR codes.These systems often lack conversational memory and cannot process mixed media content.Source: §1 · p. 1
• LURE deploys autonomous LLM agents that discover illicit channels, engage actors, and interpret visual and textual content through OCR.The system shifts LLMs from passive detection to active participation in adversarial ecosystems.Source: §1 · p. 2
• Across 98 Telegram groups, LURE identified 120 suspected providers and engaged 53, with 30 actors sustaining natural interactions and disclosing 62 payment methods.Payment disclosures included image-embedded details requiring OCR, while conversations included negotiation, upselling, and third-party redirection.Source: §1 · p. 2
• The work proposes LLM-driven engagement as a proactive strategy for exposing scam tactics and disrupting trust-based cybercrime.The authors envision agents that absorb adversarial attention, waste criminals’ time, and create uncertainty about potential victims.Source: §1 · p. 2

2 BACKGROUND

• The background frames online chat as a low-risk, adaptable entry point for scams ranging from romance fraud and phishing to impersonation and identity theft.Table 1 organizes common chat-initiated crimes by platform, audience, and use of impersonation.Source: §2 · pp. 2–3
• Visual examples illustrate how Telegram conversations and in-app functionality provide the operational setting for these activities.The supplied figures depict Telegram API utility and a nude video chat conversation.Source: Figure 1 · p. 3, Figure 2 · p. 3
• Chat-initiated crimes rely on impersonation, adaptive psychological manipulation, and scalable access to targets across platforms.Attackers use anonymity, weak monitoring, and encrypted channels to establish trust and apply pressure.Source: §2 · pp. 2–3
• Telegram’s in-app APIs expose messaging, group management, media handling, bots, and automation capabilities.Libraries such as Telethon provide broad access through Telegram’s Bot API and MTProto API.Source: §2 · p. 3

3.1 Motivation

• Rule-based defenses miss malicious intent distributed across conversations and information embedded in images.The paper identifies conversational buildup and image-embedded instructions as key blind spots.Source: §3.1 · p. 3
• LLMs paired with OCR and image recognition could interpret nuanced dialogue and visual cues while simulating victims for intelligence collection.The proposed use includes identifying deceptive intent, analyzing QR codes, and gathering attacker tactics without exposing real users.Source: §3.1 · pp. 3–4

3.2 Threat Model and Scope

• Video sex chat scams involve explicit webcam or streaming interactions that criminals may exploit for impersonation, sextortion, and blackmail.Victims may be threatened with release of recordings unless they provide money or other favors.Source: §3.2 · p. 4
• The study selects video sex chat because its trust-building and impersonation dynamics provide a controlled setting for analyzing attacker tactics and psychological manipulation.The domain is legally ambiguous, yet can serve as a gateway to more severe crimes.Source: §3.2 · p. 4

4 DESIGN OF LURE

• LURE uses a three-phase workflow covering channel discovery, target-account identification, and LLM-guided engagement with illicit service providers.The phases are designed to mimic real user behavior while scaling data collection and minimizing manual effort.Source: §4 · p. 4

4.1 Channel Discovery

• Telegram channel discovery bootstraps searches with directory bots, curated keywords, and LLM-generated synonyms, then recursively follows cross-promoted links.Telethon collects messages, media, pinned posts, and metadata from the expanding channel set for downstream filtering.Source: §4 · p. 4, §4.1 · pp. 4–5

4.2 Target Account Identification

• An LLM filters candidate channels and identifies accounts associated with illicit service promotion, while borderline or refused cases receive human review.The pipeline stores extracted data, links flagged messages to user accounts, and escalates uncertainty to annotators to reduce false positives.Source: §4 · p. 4, §4.1 · p. 5, §4.2 · p. 5

4.3 LLM-Based Engagement and Data Collection

• The engagement chatbot begins with a vague pricing inquiry, then adopts a curious casual-customer persona to elicit services, prices, payment methods, and account details.A predefined system prompt emphasizes short colloquial responses and transaction-focused conversation while maintaining a passive but responsive victim profile.Source: §4.3 · pp. 5–6
• Sensitive-word substitution and OCR help the LLM handle filtered language and images containing QR codes, payment details, or visual service menus.Conversations terminate after valid payment information is obtained, the actor disengages, or repeated model refusals persist; all interactions are manually reviewed for safety.Source: §4.3 · p. 6

5.2 Experiment Results

• 53 active providers were confirmed from 120 identified accounts across 98 Telegram groups using LURE.All 120 identified accounts were contacted, and 53 were confirmed as active illicit-service providers.Source: §5.2 · p. 6
• 30 of 53 interactions succeeded without bot identification, yielding a 56.6% success rate after excluding 15 providers that never responded.The remaining 8 failures were hypothesized to involve bot recognition, although none explicitly stated that they recognized the bot.Source: §5.2 · p. 7
• 30 of 53 confirmed providers disclosed payment methods, with Alipay_P (25.81%) and USDT (24.19%) most frequent.The analysis covered 62 disclosures because providers could report multiple payment types; disclosures came through text, QR codes, or screenshots.Source: §5.2 · pp. 7–8
• Service duration and price showed a positive but non-linear relationship, including a median above 600 Yuan for 40–44-minute sessions.The 30–34-minute interval ranged from approximately 250 to over 600 Yuan, while several intervals had only one observed price.Source: §5.2 · pp. 8–9
• Organized platforms responded more slowly but consistently, whereas individual operators showed rapid response bursts.Providers were classified using response content: platforms typically sent multiple images, while individuals usually sent one person’s image.Source: §5.2 · p. 9
• Response times remained relatively consistent across the first three interaction rounds, while the experiment was limited to five rounds.The authors suggest that these short engagements may reflect rapid targeting and turnover among cybercriminals.Source: §5.2 · pp. 9–10

Case Study 1: Redirection to a Third-Party App

• Mugua Video redirected users from an illicit video-chat interaction into an ecosystem combining adult content, gambling, and third-party payment flows.The app required account recharges and phone-number registration, while supporting Bank and Alipay payments.Source: §6 · p. 10
• The app requested camera, file-storage, and automatic APK-installation permissions, creating risks involving intimate-content blackmail, sensitive-data access, and unwanted software installation.The authors found no direct evidence that the app secretly installed malicious software.Source: §6 · p. 10
• Mugua Video also encouraged users to recruit others for rewards, linking gambling promotion with a referral-based pyramid or multi-level marketing structure.The recruitment model could produce financial benefits for higher-level participants through downstream referrals.Source: §6 · pp. 10–11, Figure 12 · p. 11

Case Study 2: Interaction with a “Bot”

• The bot interaction began with automated ordering instructions but shifted toward personalized, human-like exchanges after the initial response.The transition may reflect human takeover after customer interest or continued management by a chat-based LLM.Source: §6 · p. 11, Figure 13 · p. 11
• LLM-based scam systems could make cybercriminal interactions more personalized and convincing by adapting to victims’ tone, behavior, and emotional cues.This creates a new dimension in the defensive cat-and-mouse dynamic because automated dialogue can resemble human conversation.Source: §6 · p. 11

7.1 Board Impacts and Ultimate Objective

• LLM engagement achieved a 56% success rate, while response patterns distinguished faster individual operators from slower platforms and most scams occurred within five rounds.These findings support early detection and targeted defense strategies, with deployment balanced against cost and impact.Source: §7.1 · p. 11
• Scaling bots across illicit networks could waste criminal operators’ time, deprive them of real clients, and reduce the profitability of chat-based cybercrime.The authors envision automated interactions that simulate many virtual individuals without providing sensitive data.Source: §7.1 · p. 12
• The experiment protected participants by anonymizing collected data, deleting it after the study, avoiding third-party sharing, and not saving sensitive photos.The researchers also reported IRB approval and legal consultation.Source: §7.1 · p. 12
• Ethical concerns constrained data collection because LLM conversations could generate harmful content or expose sensitive information.The study therefore required manual monitoring and assessment before messages were sent.Source: §7.1 · p. 12

8 RELATED WORK

• Existing LLM cybercrime defenses primarily analyze linguistic patterns, contextual cues, abusive language, or code after artifacts have been produced.These applications include fraud detection, phishing and social-engineering analysis, abusive-language evaluation, and malware-defense code analysis.Source: §8 · p. 12
• LURE shifts this paradigm by deploying LLMs as real-time conversational agents that mimic human behavior while engaging cybercriminals directly.The paper presents this as the first operationalization of active LLM participation in adversarial communication ecosystems.Source: §8 · p. 12

9 CONCLUSION

• LURE sustained realistic conversations with illicit video-chat actors, often without being detected as a bot.The system’s large-scale Telegram deployment also revealed consistent behavioral patterns in scam operations.Source: §9 · p. 12
• The findings support proactive, scalable, AI-driven defenses that infiltrate and disrupt trust-based cybercrime ecosystems.The conclusion presents active conversational engagement as a new direction beyond passive detection.Source: §9 · p. 12

• At least 56% interaction success supports using LLMs to simulate victims, while deployment should balance prevention against economic losses.The authors propose calculating this trade-off to design more efficient mitigation strategies.Source: A DEFENSE STRATEGY · p. 14
• Individual operators respond quickly, whereas platform operators respond more slowly but consistently, motivating frequent monitoring for individuals and long-term strategies for platforms.Source: A DEFENSE STRATEGY · p. 14
• Most criminals respond quickly and almost all scams are completed within five rounds, making early-stage detection a resource-efficient defense priority.The proposed strategy is to identify and stop high-risk conversations before they escalate.Source: A DEFENSE STRATEGY · p. 14