Zorina works with global companies to find solutions that speed up operations and enhance processes using Artificial Intelligence and Machine Learning. Zorina is also an Adjunct Professor at Georgetown University SCS, as a creator and instructor for the “AI for Leaders” courses within the Project Management Masters program.

Zorina is involved in AI for Good initiatives and working with non-profit organizations to address major social and environmental challenges using AI/ML. She also volunteers with the Zonta organization and as the Chair of the Artificial Intelligence Committee at AnitaB.org.

Network Centrality, Adversarial Trading, and Crisis Deterrence: Evidence from a Multi-Agent Market Crisis Simulation

What would happen if AI agents were given real autonomy to trade during a financial crisis?
This paper takes a first step toward answering that question. We investigate whether graph-
aware defensive market-making agents can neutralise the destabilising effects of adversarial
traders who exploit a financial contagion network’s topology, using a multi-agent simulation
platform calibrated to the March 2023 Silicon Valley Bank (SVB) crisis. By observing how
AI agents — adversarial, defensive, and neutral — behave autonomously across 200 Monte
Carlo iterations, we generate empirical evidence about their emergent strategies, failure
modes, and mutual deterrence dynamics. We design two experiments: a competition among
three graph exploiters against a single centrality-weighted defender, and a confrontation in
which a single graph exploiter faces a coalition of three specialist defenders. Each scenario
runs 100 stochastic iterations of 100 discrete time steps. Our results show that the three-
defender coalition reduces mean realised volatility by 90.6% (from 0.228% to 0.021%;
Welch’s t = 10.00, p < 0.001, Cohen’s d = 1.44) and mean price drawdown by 86.7%
(Welch’s t = 7.61, p < 0.001, d = 1.11). Market activity collapses by 97.6% (from 19.0 to 0.45
mean trades per iteration), driven by a 51.6% widening of bid-ask spreads. All three defender
strategies record zero realised PnL, revealing a deterrence mechanism: defenders neutralise
adversaries not through active counter-trading but by raising the cost of adversarial execution
beyond viability. We term this the deterrence paradox: the most effective defence is one that
is never invoked. Critically, all adversarial agents are loss-making even without dedicated
defence, demonstrating that graph-aware exploitation is self-defeating under realistic
transaction costs — an important behavioural finding for AI safety in financial markets.
These results contribute to the project of understanding AI agent behaviour well enough to
deploy such agents safely, transparently, and with predictable outcomes in real financial
crises.