---
title: Agentic AI Systems and Financial Stability, From Model Risk to Systemic Risk
url: https://www.ml-quant.com/papers/arxiv/2610.08806/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-10-09
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: arXiv:2610.08806
source_url: https://arxiv.org/abs/2610.08806
featured: 2026-10-09
citations: unknown
topic: Risk, Credit & Banking
---


# Agentic AI Systems and Financial Stability, From Model Risk to Systemic Risk

The study shows that agentic AI systems sharing a foundation model create non-diversifiable common exposure whose systemic risk floor does not shrink as the fleet grows, shifting concern from individual model risk to population-level contagion.

- Source: https://arxiv.org/abs/2610.08806
- Identifier: arXiv:2610.08806
- Released: 2026-10-08
- First featured: Quant Letter No. 134 (2026-10-09): https://www.ml-quant.com/issues/2026-10-09/
- Citations (Semantic Scholar): not tracked
- Published in: not yet
- Topic: Risk, Credit & Banking
- Authors: Sriram Nagaraj, Seung Jung Lee

## Abstract (arXiv, CC0)

Financial stability rests on the premise that distress is largely idiosyncratic and therefore diversifiable: when one institution errs, the rest of the system absorbs the shock. That premise fails when many decision-makers depend on the same infrastructure. Agentic AI systems, which take consequential actions rather than only emitting predictions, are becoming such a substrate, and the binding concern shifts from the model risk of a single deployment to the systemic risk of the population. We develop this account in six mathematical settings. Chapter 1 recasts the PD/LGD/EAD decomposition of expected loss as an expected-harm identity and introduces a set-valued containment-risk measure; under a lever-assignment axiom that review cannot intercept an irreversible action, no non-preventive control satisfies a coherent tail constraint. Chapter 2 lifts this to a fleet sharing a foundation model: the shared model is a non-diversifiable common exposure whose expected-shortfall floor holds however large the fleet grows, and a percolation threshold governs contagion. Chapter 3 recasts these dynamics as a marked, Hawkes-excited jump diffusion, identifying the robust stress problem with a time-consistent entropic risk measure. Chapter 4 treats runtime guardrails as partially observed stochastic control, giving an observability trichotomy, a detection floor, and, under commit exogeneity, a runtime-impossibility corollary. Chapter 5 makes the adversary a player, yielding an underinvestment wedge proportional to systemic reach. Chapter 6 makes capability a state variable in an arms race. One claim runs through all six: the systematic component of agentic risk cannot be diversified, detected away, or reversed, and only ex-ante structural prevention moves it.

## Related

- [AI Agents in Financial Markets: Architecture, Applications, and Systemic Implications](https://www.ml-quant.com/papers/arxiv/2603.13942/): Recent AI advancements are enhancing financial automation by creating integrated systems that use autonomous agents for better decision-making and processing, highlighting the need for effective agent governance.
- [Systemic Risk in the European Insurance Sector](https://www.ml-quant.com/papers/arxiv/2505.02635/): The study investigates the relationship between the European insurance sector and financial markets, finding that the insurance market contributes to systemic risk, especially during financial crises.
- [Systemic risk mitigation in supply chains through network rewiring](https://www.ml-quant.com/papers/arxiv/2504.12955/): Research shows that strategically rearranging supplier-customer connections can significantly lower supply chain risk by 16-50% without affecting production.
- [A data-driven econo-financial stress-testing framework to estimate the effect of supply chain networks on financial systemic risk](https://www.ml-quant.com/papers/arxiv/2502.17044/): The research investigates how systemic risk in production networks can lead to financial systemic risk through supply chain contagion, and proposes a financial stress-testing framework that combines supply chain and interbank network layers.
- [A novel k-generation propagation model for cyber risk and its application to cyber insurance](https://www.ml-quant.com/papers/arxiv/2408.14151/): The research proposes a new model for calculating aggregate losses in cyber insurance pricing, considering the origin contagion location and varying security levels in a network.
- [Global balance and systemic risk in financial correlation networks](https://www.ml-quant.com/papers/arxiv/2407.14272/): The study proves that the global balance index of financial correlation networks can effectively measure systemic risk, as confirmed by its application to real financial data.