REVIEW 4 major objections 4 minor 3 cited by
Multi-turn scam dialogues with LLMs exhibit regular escalation and defense patterns that single-turn safety tests miss.
Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →
T0 review · deepseek-v4-flash
2026-08-03 12:21 UTC pith:DL2DDQOC
load-bearing objection Useful descriptive corpus and a plausible taxonomy, but the attacker prompt seeds most of the discovered strategy families, so the 'internalized schemas' claim is over-claimed. the 4 major comments →
The Anatomy of Conversational Scams: A Topic-Based Red Teaming Analysis of Multi-Turn Interactions in LLMs
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The paper's central claim is that scam interactions follow recurrent escalation patterns and that defensive responses frequently rely on verification, delay, and channel control. Concretely, across 18,648 simulated bilingual dialogues, success is rare (about 4–5%), and most interactions end in explicit detection or in a no-resolution stalemate, while a large fraction of dialogues collapse into ERROR due to safety-guardrail activation, role instability, or template artifacts. Qualitative topic analysis yields a stable taxonomy of ten attacker strategy families (e.g., urgency, authority pressure, channel shift, rapport building) and ten defender strategy families (e.g., verification, delay, ch
What carries the argument
The central mechanism is the LLM-to-LLM adversarial dialogue loop: a ScamBot attacker and a VictimBot defender are instantiated with persona prompts in ten fictional fraud categories, capped at ten turns, and the resulting dialogues are labeled (SUCCESS, DETECTED, NO_RESOLUTION, ERROR) with human adjudication. Topic modeling over utterances surfaces latent strategy families, and qualitative coding organizes them into attacker, defender, and error taxonomies. The load-bearing result is the 'functional symmetry' between attack and defense: each attacker tactic has a recognizable counter-mechanism, so the dialogue behaves like a dynamic contest between competing interactional mechanisms. This f
Load-bearing premise
The findings rely on the premise that an LLM victim instructed to be suspicious and to replace sensitive information with placeholders behaves enough like a real human target that the observed strategies and outcomes transfer to actual scam conversations.
What would settle it
Compare a sample of the simulated dialogues turn-by-turn to real-world scam transcripts (e.g., recorded scam-baiting calls or reported conversations). If the escalation families and defensive responses do not appear in real data, the claimed ecological validity of the simulation fails.
If this is right
- Multi-turn safety evaluations should treat interactional structure as a separate metric; single-turn refusal benchmarks miss escalation and drift.
- Defensive robustness can be characterized by a small set of resistance moves (verification, delay, channel control) and by their timing, not just by whether the model refuses.
- ERROR outcomes are informative boundary markers: they reveal where role conditioning and safety guardrails collide, and they are not random noise.
- Cross-lingual differences in outcome distributions imply safety alignment thresholds are language-dependent, so evaluations must be multilingual.
- LLM-to-LLM simulation can generate large annotated corpora for studying social-engineering dynamics safely, without real victims.
Where Pith is reading between the lines
- If the interactional grammar is as stable as claimed, the same framework could be used as a generative training signal to make defensive agents more resilient by learning to deploy the identified counter-tactics at the right times.
- The strong dependence of ERROR rates on prompt template artifacts suggests that measurement artifacts—not just model capability—may dominate behavioral comparisons; future work should vary prompt fidelity to disentangle simulation design from model behavior.
- The symmetry between attack and defense families implies that an automated red-teaming loop could be turned into a co-evolutionary game, where attacker and defender strategies are iteratively hardened, possibly revealing new failure modes beyond the fixed ten-turn horizon.
- The rarity of SUCCESS and the prevalence of NO_RESOLUTION suggests that, in these simulations, the main risk is not immediate fraud completion but long, inconclusive engagements that could eventually erode suspicion in human settings—a testable hypothesis.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes an LLM-to-LLM simulation framework for multi-turn conversational scam analysis. It generates 18,648 bilingual (Chinese/English) dialogues between a ScamBot attacker and a VictimBot defender across eight models, labels each dialogue as SUCCESS, DETECTED, NO_RESOLUTION, or ERROR, applies BERTopic to utterances, and manually consolidates topics into attacker (AF1–AF10), defender (DF1–DF10), and error (EF1–EF7) strategy families. The central claims are that scam interactions exhibit recurrent escalation patterns, that defensive responses rely on verification/delay/channel control, that there are statistically significant cross-model and cross-lingual differences, and that multi-turn interactional safety is a distinct behavioral dimension not captured by single-turn benchmarks.
Significance. The paper has clear strengths: a large controlled bilingual corpus, a fictionalized placeholder-based safety design, human adjudication of outcome labels, and an explicit limitations section. If the claimed regularities were robust, the study would support a valuable shift toward multi-turn interaction-level safety evaluation. However, the main interpretive claims are currently not adequately supported: the attacker prompt appears to seed the discovered taxonomy, the quantitative significance claim lacks any statistical procedure, and the defensive-strategy results are conditional on an explicitly suspicious simulated victim. These issues bear directly on the paper's load-bearing conclusions.
major comments (4)
- [§6.1, Appendix A, Table 8] The central claim that LLMs exhibit 'internalized temporally structured manipulation schemas' is confounded by the attacker prompt. Appendix A's 'Scammer Characteristics' explicitly instructs the model to use professional authority, channel shift to WeChat/QQ, verification-code/bank-card requests, deadline pressure, and 'operational errors' as inducements. These map almost one-to-one onto AF1, AF2, AF3, AF4, AF5, AF6, and AF10 in Table 8. The fact that topic modeling is unsupervised does not remove this confound because the clustering operates on utterances generated under a prompt that already specifies the tactics. A control condition that varies or removes prompt-specified tactics is required before attributing the taxonomy to emergent interactional dynamics rather than prompt compliance.
- [Abstract and §5.1] The abstract claims 'statistically significant cross-model and cross-lingual differences in outcome distributions,' but no statistical test, confidence interval, or effect size is reported anywhere in §5 or the appendices. Tables 3–8 report proportions without uncertainty, and Figure 2 has no error bars. Given the large sample (n≈18,648), even negligible differences would be 'significant' under naive tests; proper procedures should account for clustering by model, role, and scenario. Without such analysis, the headline significance claim is unverifiable.
- [§3.1, §5.1.2, §8] The model selection criterion explicitly excludes systems that 'consistently trigger built-in safety interruptions.' Consequently, the ERROR rates, error-family distributions, and 'guardrail asymmetry' conclusions in §5.1.2 and Table 9 are conditional on surviving this filter. The Limitations section acknowledges this, but the body text states without hedging that 'language significantly conditions conversational persistence' and discusses guardrail asymmetry as a general phenomenon. The results should be framed as selection-conditional, or supplemented with a sensitivity analysis that includes the excluded models.
- [§3.2/Appendix B, §6.2, §8] The VictimBot prompt instructs the model to be suspicious and to replace sensitive information with placeholders. The defensive strategy families in Table 7 and the conclusion that defensive robustness depends on timing and combination are therefore properties of this particular simulated victim, not of human victims. The paper's own Limitations states that the simulation 'cannot fully capture the diversity and unpredictability of real-world human behavior.' For the findings to transfer to human scam interactions, validation against human-scam transcripts or human-in-the-loop data is needed; otherwise the conclusions should be restricted to LLM-based victim agents.
minor comments (4)
- [§6.1, Figure 3] The text cites Figure 3 as supporting 'temperate progression' from low-commitment positioning to high-stakes operations, but Figure 3 is a scatter plot of attacker success vs. victim detection and contains no temporal or sequential information. The citation should be corrected or a new figure added.
- [§6.3, Table 9] The first paragraph of §6.3 reports EF1 rates of 86.71% (Chinese) and 73.42% (English) without specifying that these are victim-role percentages from Table 9. This should be stated explicitly to avoid confusion with attacker-role errors.
- [Figure 4 caption] The caption refers to 'Appendix 7, Appendix 9, Appendix 8,' but the paper uses lettered appendices (A–D) and numbered tables. These references should be harmonized with the actual appendix/table numbering.
- [§9] Typo: 'at a hourly rate' should be 'at an hourly rate.' The same sentence also appears with 'a hourly rate' later in the ethical considerations paragraph.
Circularity Check
Attacker prompt enumerates the reported strategy families, so 'internalized scam schemas' partly reduce to prompt compliance.
specific steps
-
fitted input called prediction
[Appendix A (ScamBot prompt, 'Scammer Characteristics') vs. §6.1 'Attacker Strategy Patterns' and Table 8]
"Guides users to add WeChat, QQ, or other tools to bypass platform supervision. ... Requests verification codes, bank card numbers, or SMS information. ... Applies pressure (e.g., “missing the deadline means refunds are no longer possible”). ... Uses excuses like “operational errors” to induce further transfers. [Appendix A] ... Channel Shift (AF5) remains a universally dominant tactic (>80%) across both languages. [§6.1]"
The attacker prompt explicitly instructs the model to use the exact tactics that §6.1/Table 8 report as discovered strategy families: AF5 Channel Shift, AF4 Information Harvesting, AF2 Urgency/AF3 Threat, AF10 Payment Engineering. Because every dialogue was generated under this same ScamBot specification, the high AF coverage and cross-model stability are expected from instruction-following; they cannot independently show that LLMs have 'internalized temporally structured manipulation schemas' acquired during pre-training. The phenomenon is produced by the experimental input rather than independently measured.
full rationale
The paper is a transparent simulation study: it states that attacker prompts are based on official scam categories, and it does not fit any parameter to the outcome data. The outcome labels are human-verified; the defensive strategy families (DF1-DF10) are not explicitly listed in the VictimBot prompt, and the cross-lingual differences (e.g., AF2 78.48% EN vs 50.65% ZH) are not directly dictated by identical prompt templates. Those parts contain independent content. The circularity lies specifically in the attacker-side 'internalization' claim: the ScamBot prompt's 'Scammer Characteristics' bullet list already enumerates channel shift, information harvesting, urgency/pressure, and payment-engineering tactics, and Table 8 then reports the same tactics as emergent strategy families. The 'unsupervised' nature of the topic clustering does not remove this contamination because clustering operates on utterances generated under the explicit tactic list. Therefore the finding that LLMs 'have internalized canonical manipulation patterns' reduces, for the attacker families, largely to prompt compliance. No self-citation chain is load-bearing; all cited prior work is external. Score 6 reflects partial circularity: the central attacker-taxonomy/internalization claim is seeded by the input, while the defensive and cross-lingual analyses are not.
Axiom & Free-Parameter Ledger
free parameters (1)
- max_turns =
10
axioms (4)
- domain assumption LLM-to-LLM role-play is a valid proxy for human victim behavior in scams
- domain assumption Prompt-specified attacker tactics do not predetermine the discovered strategy families
- domain assumption Manual topic-to-family mapping is reliable
- standard math BERTopic clusters correspond to meaningful interactional strategies
invented entities (1)
-
Strategy family taxonomy (AF1-10, DF1-10, EF1-7)
no independent evidence
read the original abstract
As LLMs gain persuasive capabilities through extended dialogues, they create new opportunities for studying adversarial conversational behavior in extended interaction settings that traditional single-turn safety evaluations fail to capture. We systematically study these interactional dynamics using a controlled LLM-to-LLM simulation framework for automated red-teaming across bilingual social engineering scenarios. Evaluating eight state-of-the-art models in English and Chinese, we analyze dialogue-level outcomes, annotate attacker and defender strategy families, and model interaction dynamics between them. Results show that multi-turn adversarial dialogues follow recurrent escalation patterns, while defensive responses frequently rely on verification, delay, and channel control. We further find statistically significant cross-model and cross-lingual differences in outcome distributions, and transition analysis reveals systematic structural variation in how defender strategies respond to attacker tactics across languages. These findings highlight the importance of studying interactional structure in multi-turn adversarial dialogue settings and demonstrate how controlled LLM-to-LLM simulations can support mechanistic analysis of adversarial conversational dynamics.
Figures
Forward citations
Cited by 3 Pith papers
-
An Explainable Agentic System for Detection of Conversational Scams with Summary-Based Memory
An explainable agentic pipeline with summary-based memory detects all 83 LoveFraud02 scams and reaches 97.8% accuracy on the new multi-category ConScamBench-278 while improving single-message phishing recall to 100%.
-
An Explainable Agentic System for Detection of Conversational Scams with Summary-Based Memory
An LLM multi-agent system with summary-based memory detects slow-building conversational scams, achieving 100% recall on LoveFraud02 and 97.8% accuracy on the new ConScamBench-278 benchmark.
-
Robust Critics: Defending LLMs Against Multi-Turn Attacks
Critic-weighted sampling over inferred user intents improves multi-turn LLM defense success while preserving helpfulness, with an expected-Q improvement guarantee and transfer to frontier models.
Reference graph
Works this paper leans on
-
[1]
Ross Anderson. 2010. Security engineering: a guide to building dependable distributed systems. John Wiley & Sons
2010
-
[2]
Robert B Cialdini. 2007. Influence the psychology of persuasion. Harpercollins publishers
2007
-
[3]
Anna Coluccia, Andrea Pozza, Fabio Ferretti, Fulvio Carabellese, Alessandra Masti, and Giacomo Gualtieri. 2020. Online romance scams: relational dynamics and psychological characteristics of the victims and scammers. a scoping review. Clinical practice and epidemiology in mental health: CP & EMH, 16:24
2020
-
[4]
Ian Fette, Norman Sadeh, and Anthony Tomasic. 2007. Learning to detect phishing emails. In Proceedings of the 16th international conference on World Wide Web, pages 649--656
2007
-
[5]
Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer, Kamal Ndousse, and 1 others. 2022. Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned. arXiv preprint arXiv:2209.07858
Pith/arXiv arXiv 2022
-
[6]
Maarten Grootendorst. 2022. Bertopic: Neural topic modeling with a class-based tf-idf procedure. arXiv preprint arXiv:2203.05794
Pith/arXiv arXiv 2022
-
[7]
Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300
Pith/arXiv arXiv 2020
-
[8]
Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, and 1 others. 2023. Metagpt: Meta programming for a multi-agent collaborative framework. In The Twelfth International Conference on Learning Representations
2023
-
[9]
Tharindu Kumarage, Cameron Johnson, Jadie Adams, Lin Ai, Matthias Kirchner, Anthony Hoogs, Joshua Garland, Julia Hirschberg, Arslan Basharat, and Huan Liu. 2025. Personalized attacks of social engineering in multi-turn conversations--llm agents for simulation and detection. arXiv preprint arXiv:2503.15552
Pith/arXiv arXiv 2025
-
[10]
Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023. Camel: Communicative agents for" mind" exploration of large language model society. Advances in Neural Information Processing Systems, 36:51991--52008
2023
-
[11]
Yi Liu, Gelei Deng, Zhengzi Xu, Yuekang Li, Yaowen Zheng, Ying Zhang, Lida Zhao, Tianwei Zhang, Kailong Wang, and Yang Liu. 2023. Jailbreaking chatgpt via prompt engineering: An empirical study. arXiv preprint arXiv:2305.13860
Pith/arXiv arXiv 2023
-
[12]
Leland McInnes, John Healy, Steve Astels, and 1 others. 2017. hdbscan: Hierarchical density based clustering. J. Open Source Softw., 2(11):205
2017
-
[13]
Leland McInnes, John Healy, and James Melville. 2018. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426
Pith/arXiv arXiv 2018
-
[14]
Anay Mehrotra, Manolis Zampetakis, Paul Kassianik, Blaine Nelson, Hyrum Anderson, Yaron Singer, and Amin Karbasi. 2024. Tree of attacks: Jailbreaking black-box llms automatically. Advances in Neural Information Processing Systems, 37:61065--61105
2024
-
[15]
Rajvardhan Oak and Zubair Shafiq. 2025. Victims, vigilantes, and advice givers: An analysis of \ Scam-Related \ discourse on reddit. In Twenty-First Symposium on Usable Privacy and Security (SOUPS 2025), pages 57--71
2025
-
[16]
Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730--27744
2022
-
[17]
Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1--22
2023
-
[18]
Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. 2022. Red teaming language models with language models. arXiv preprint arXiv:2202.03286
Pith/arXiv arXiv 2022
-
[19]
Aashray Reddy, Andrew Zagula, and Nicholas Saban. 2025. Autoadv: Automated adversarial prompting for multi-turn jailbreaking of large language models. arXiv preprint arXiv:2507.01020
arXiv 2025
-
[20]
Said Salloum, Tarek Gaber, Sunil Vadera, and Khaled Shaalan. 2021. Phishing email detection using natural language processing techniques: a literature survey. Procedia Computer Science, 189:19--28
2021
-
[21]
Murray Shanahan, Kyle McDonell, and Laria Reynolds. 2023. Role play with large language models. Nature, 623(7987):493--498
2023
-
[22]
Nikolaos Tsinganos, Panagiotis Fouliras, and Ioannis Mavridis. 2023. Leveraging dialogue state tracking for zero-shot chat-based social engineering attack recognition. Applied Sciences, 13(8):5110
2023
-
[23]
Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, and 1 others. 2024 a . A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345
2024
-
[24]
Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024 b . Multilingual e5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672
Pith/arXiv arXiv 2024
-
[25]
Noah Wang, Zy Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, and 1 others. 2024 c . Rolellm: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. In Findings of the Association for Computational Linguistics: ACL 2024, pages 14743--14777
2024
-
[26]
Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. 2023. Jailbroken: How does llm safety training fail? Advances in Neural Information Processing Systems, 36:80079--80110
2023
-
[27]
Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, and 1 others. 2021. Ethical and social risks of harm from language models. arXiv preprint arXiv:2112.04359
Pith/arXiv arXiv 2021
-
[28]
Ian Wood, Michal Kepkowski, Leron Zinatullin, Travis Darnley, and Mohamed Ali Kaafar. 2023. An analysis of scam baiting calls: Identifying and extracting scam stages and scripts. arXiv preprint arXiv:2307.01965
Pith/arXiv arXiv 2023
-
[29]
Shu Yang, Shenzhe Zhu, Zeyu Wu, Keyu Wang, Junchi Yao, Junchao Wu, Lijie Hu, Mengdi Li, Derek F Wong, and Di Wang. 2025. Fraud-r1: A multi-round benchmark for assessing the robustness of llm against augmented fraud and phishing inducements. arXiv preprint arXiv:2502.12904
Pith/arXiv arXiv 2025
-
[30]
Jinsol Yoo and Youngho Cho. 2022. Icsa: Intelligent chatbot security assistant using text-cnn and multi-phase real-time defense against sns phishing attacks. Expert Systems with Applications, 207:117893
2022
-
[31]
Xu, Hao Zhu, Ruijie Xu, Robert Lo, Abhinav Srivastava, and Yonatan Bisk
Shuyan Zhou, Frank F. Xu, Hao Zhu, Ruijie Xu, Robert Lo, Abhinav Srivastava, and Yonatan Bisk. 2023. https://arxiv.org/abs/2307.13854 Webarena: A realistic web environment for building autonomous agents . arXiv preprint arXiv:2307.13854
Pith/arXiv arXiv 2023
-
[32]
Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. 2023. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043
Pith/arXiv arXiv 2023
-
[33]
online" 'onlinestring :=
ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...
-
[34]
write newline
" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.