Elamur Research⌕
☰
home / papers / 2026 / VishBox v2: A Multi-Agent System for Adaptive Voice Phishing Simulation

VishBox v2: A Multi-Agent System for Adaptive Voice Phishing Simulation

Paper · 2026 · Sungmi Park, Daon Choi, Yoonmo Yang et al.

Summary

Real voice-phishing call transcripts are rarely available to researchers, and substitutes such as static scenarios or scattered public cases do not show how attacker tactics and victim psychology shift from round to round. VishBox v2 addresses this gap with a multi-agent simulator:

  • Main Agent: plans, coordinates and scores each round.
  • Dialogue Agent: connected through MCP, it generates attacker turns with GPT-4o-mini and victim turns with Gemini-2.5-flash-lite.
  • Tactic Search Agent: retrieves newer scam techniques from the web.

Attacker behaviour is constrained by police-derived procedural codes (proc_code) and PPSE persuasion principles. Victim emotions feed an HMM that infers vulnerability states V1–V3. Focusing on South Korean law-enforcement impersonation scams, the authors produced 181 cases (571 rounds) across three susceptibility personas. Analyses, including police-expert assessment, suggest that the dialogues follow documented crime-script stages and that emotion sequences align with inferred vulnerability. They also show that web-mined tactics can disrupt procedural flow when introduced too early. The system offers a testbed that avoids private transcripts and supports anti-fraud training, detector stress-testing and case analysis.

Key points

  • Authorship: the first authors are Daon Choi and Yoonmo Yang (equal contribution), with co-authors from Hallym University and the Korean National Police Agency. Sungmi Park is the corresponding (last) author, so the library citation "Sungmi Park et al." should be corrected to "Choi, Yang et al."
  • Dataset: 181 multi-round cases (571 rounds) built from three personas (high, medium and low susceptibility). The personas vary digital financial literacy and Big Five (OCEAN) traits.
  • Risk scoring: round-level risk falls into four bins: Low (0–24), Medium (25–49), High (50–74) and Critical (75–100). A case ends after five rounds or when it reaches Critical risk.
  • Procedural patterns: the most frequent attacker proc_code 3-gram is 1-1→2-1→2-2 (identity check → affiliation claim → purpose statement), with 195 occurrences (4.27%).
  • Emotions and web search: persistently defensive emotion patterns (e.g., A→A→A) are linked to a low success rate (15.4%). Rounds using web search succeeded far less often than rounds without it (10.98% vs 50.37%). The authors note that this comparison is confounded, because search is triggered after repeated failures.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

Also on the byline: Hallym University · Korean National Police Agency.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksInformation Systems · 1,437 works

Cite

Cite as Sungmi Park, Daon Choi, Yoonmo Yang, Hong Yunyi and Heedou Kim. 2026. VishBox v2: A Multi-Agent System for Adaptive Voice Phishing Simulation. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-industry.145

@inproceedings{park2026vishbox,
  title = {{VishBox v2: A Multi-Agent System for Adaptive Voice Phishing Simulation}},
  author = {Sungmi Park, Daon Choi, Yoonmo Yang, Hong Yunyi and Heedou Kim},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-industry.145},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.145}
}

References

12 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Lin Ai, Tharindu Sandaruwan Kumarage, Amrita Bhattacharjee, Zizhou Liu, Zheng Hui, Michael S. Davinroy, James Cook, Laura Cassani, Kirill Trapeznikov, Matthias Kirchner, Arslan Basharat, Anthony Hoogs, Joshua Garland, Huan Liu, and Julia Hirschberg. 2024. Defending against social engineering attacks… find
  2. Anthropic. 2024. Introducing the model context protocol. https://www.anthropic.com/news/ model-context-protocol. ↗
  3. Federal Bureau of Investigation. 2022. 2021 Internet Crime Report. Technical report, Internet Crime Complaint Center (IC3). [Online]. Available: https://www.ic3.gov/Media/ PDF/AnnualReport/2021_IC3Report.pdf. ↗
  4. Ana Ferreira, Lynne Coventry, and Gabriele Lenzini. 2015. Principles of persuasion in social engineering and their use in phishing. In Proceedings of the Third International Conference on Human Aspects of Information Security, Privacy, and Trust - Volume 9190, page 36–47, Berlin, Heidelberg. Springe… find
  5. O. P. John, L. P. Naumann, and C. J. Soto. 2008. Paradigm shift to the integrative big five trait taxonomy: History, measurement, and conceptual issues. In O. P. John, R. W. Robins, and L. A. Pervin, editors, Handbook of Personality: Theory and Research, 3rd edition, pages 114–158. Guilford Press, N… find
  6. Heedou Kim, Changsik Kim, Sanghwa Shin, and Jaewoo Kang. 2026. Scriptmind: Crime script inference and cognitive evaluation for llm-based social engineering scam detection system. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026… find
  7. Symposium on Digital Forensics and Security (IS-DFS). find
  8. V. Nyassi, F. Tchakounté, B. Yenké, D. Danga, M. Ngo-ran, and J. Fendji. 2024. Emoti-shing: Detecting vishing attacks by learning emotion dynamics through hidden markov models. Journal of Intelligent Learning Systems and Applications, 16:274– 315. find
  9. Andrew T. Rozema and James C. Davis. 2026. Anti-phishing training (still) does not work: A reproduction of phishing training inefficacy grounded in the nist phish scale. In Proceedings of the ACM Web Conference 2026, WWW ’26, page 3147–3158, New York, NY, USA. Association for Computing Machinery. find
  10. Dawn M. Sarno, Maggie W. Harris, and Jeffrey Black. 2023. Which phish is captured in the net? understanding phishing susceptibility and individual differences. 37(4):789–803. find
  11. Andreas Triantafyllopoulos, Anika A. Spiesberger, Iosif Tsangko, Xin Jing, Verena Distler, Felix Dietz, Florian Alt, and Björn W. Schuller. 2025. Vishing: Detecting social engineering in spoken communication — a first survey & urgent roadmap to address an emerging societal challenge. Computer Speech… find
  12. Yoonmo Yang, Daon Choi, Yunyi Hong, Jee-Won Park, Jae-Yong Yu, Hee-Dou Kim, and Sungmi Park. 2026. VishBox: An AI-Agent-Based Adaptive Voice Phishing Simulation Framework for Cybersecurity Education. IEEE Access, 14:39672–39686. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/park-2026-vishbox-v2-multi-agent-system/