Imagine you run a restaurant kitchen during a dinner rush, and someone starts sending in fake orders — plausible-looking tickets for tables that don't exist. Every fake ticket burns cook time, ingredients, and attention. Now imagine you could flood the scammer's own ordering system with fake cancellations and corrections, tying up their phone line so thoroughly that real orders barely get through. That's the core mechanism here: autonomous scambaiting agents that impersonate victims, waste scammer resources, and extract intelligence (bank accounts, crypto wallets) that can be reported for takedown. The committed claim: there exists a formal, game-theoretic threshold of baiting-agent prevalence beyond which scammer operations become unprofitable, and in the resulting arms race, defenders investing in mimicry quality gain ground faster than scammers investing in detection capability. This is not an empirical demonstration — it's a mathematical framework paper. No bots were deployed. The result is a set of analytical findings about equilibrium conditions, not a field trial. The model treats fraud as an adversarial ecosystem with two populations of autonomous agents — scamming and scambaiting — competing over resource discovery, preservation, and exploitation. Scammers maintain fraudulent infrastructure (mule accounts, phishing domains) that has a finite lifespan before detection. Baiting agents extend engagement duration to extract actionable intelligence while burning scammer time and resources. The key structural insight is asymmetric: improving a baiting agent's victim-mimicry quality imposes costs on ALL scammer interactions (because scammers must now spend more effort screening every contact), while improving scammer detection only helps against baiting agents specifically. This asymmetry is what tilts the arms race. The architectural family is classical game theory and adversarial modeling — think mechanism design and population dynamics, not deep learning. The paper uses formal utility functions, belief-adaptation models, and threshold analysis rather than training neural networks. This matters because the results are parametric: they hold under stated assumptions about cost structures and agent capabilities, not under empirical conditions that might shift. The framework borrows from security economics (Anderson, Herley) and computational deception research. Integrity is the weakest link. The validation is entirely analytical — the authors are grading their own homework via their own formal model. No simulation was run against real scammer behavior, no empirical data on actual scambaiting deployments was used to calibrate parameters, and no sensitivity analysis is reported for the threshold values. The finding that 'scammer profitability turns negative once baiting-agent prevalence exceeds certain thresholds' is a mathematical consequence of the model's assumptions, not an observed phenomenon. The baseline comparison is implicit: prior scambaiting work (Apate, Li et al. 2024) is discussed qualitatively, but no quantitative comparison of effectiveness metrics is presented. The milestone question is concrete but distant. The paper identifies the critical variable — baiting-agent prevalence threshold — but does not quantify it in deployable terms (e.g., 'you need X thousand baiting agents per Y million scam attempts'). The next real unlock would be an empirical calibration: deploying autonomous scambaiting agents at scale against live scam operations and measuring whether profitability inversion actually occurs at predicted thresholds. That experiment was not run, most likely because it requires partnerships with telecoms, banks, or law enforcement that an academic team cannot easily arrange — and possibly because real-world scammer adaptation would complicate the clean analytical results. This is a first-principles framework paper, not an engineering result. Its value is in formalizing intuitions that the scambaiting community has operated on informally: that wasting scammer time works, that scale matters, and that defenders have a structural advantage in the mimicry-vs-detection arms race. Whether those formal results survive contact with the messy reality of actual fraud ecosystems is the open question.
