The banking sector now relies heavily on a small set of shared AI vendors that provide fraud screening, credit decisioning, anti‑money‑laundering triage, customer analytics, and internal decision support. This paper investigates how a breach in one of these vendors can spread through operational, informational, and financial linkages, eventually manifesting as a classic banking crisis. We construct a four‑layer heterogeneous network:
- AI vendor layer;
- Financial institution layer;
- Interbank exposure layer;
- Customer account layer.
On this network we propose CFC‑Prop, a stochastic epidemic‑and‑clearing model. The core mechanism simulates infection probability on each vendor‑bank service edge; once infected, a bank suffers an asset shock $\Delta A = \beta \times L$, where $\beta$ denotes attack intensity and $L$ the dependency weight on the compromised vendor. The shock then propagates through the interbank exposure network via a clearing process solved by the Eisenberg‑Noe fixed‑point equation for the payment vector $p$.
We evaluate the model on a synthetic dataset comprising 60 vendors, 220 banks, roughly 2,500 vendor‑bank service edges, and 1,400 interbank exposures. Results show that CFC‑Prop reproduces heavy‑tailed loss distributions and exhibits a sharp dependence on patch latency, consistent with prior cyber‑financial observations.
For early warning, we train a graph neural network called CFC‑GNN that ingests vendor‑side incident telemetry and graph structure to output a cascading‑risk score. Compared with four baselines, CFC‑GNN achieves AUROC 0.82 and AUPRC 0.60 while keeping calibration errors bounded.
All code, synthetic data, and reproducible scripts are released publicly, enabling supervisors and researchers to apply the framework directly. The findings argue that concentration among AI vendors constitutes a first‑order financial‑stability problem, urging regulators to adopt quantitative assessment and mitigation strategies.
Review: By rigorously linking cyber‑risk propagation with systemic financial contagion, the paper provides a valuable quantitative tool for supervision, bridging a critical gap between AI security and macro‑prudential oversight.