AI Agents Could Trigger Bank Runs, Warns Apollo Economist

What happened

Torsten Slok, Chief Economist at Apollo Global Management, has issued a warning regarding the potential for autonomous AI agents to precipitate a bank run. This concern stems from the rapid, decentralised decision-making capabilities of these AI systems, which could collectively withdraw funds from financial institutions if a perceived risk or negative signal were to emerge. Such a scenario, dubbed an "agentic bank run," highlights a novel systemic risk introduced by advanced AI applications in the financial sector.

Why it matters

This development underscores a critical, emerging challenge for AI governance and risk management within the financial industry. The proliferation of AI agents capable of independent action and rapid transaction execution introduces complex systemic vulnerabilities that traditional regulatory frameworks may not adequately address. For business leaders and policymakers, it necessitates a proactive re-evaluation of financial stability protocols, ethical AI deployment in critical infrastructure, and the development of robust oversight mechanisms to mitigate unforeseen cascade effects. The potential for AI-driven collective action, even if unintentional, demands urgent attention to prevent financial instability and maintain public trust in AI-powered systems.

The Institute take

The discussion around "agentic bank runs" reveals a fundamental oversight in much of the current AI risk discourse: the focus often remains on individual AI failures rather than the aggregate, systemic risks posed by numerous autonomous agents interacting within complex environments. While the immediate threat of AI-driven financial instability is concerning, the deeper lesson lies in understanding the emergent behaviours of interconnected AI systems.

Organisations deploying AI agents, particularly in high-stakes sectors like finance, must move beyond mere model-level risk assessments. A truly responsible approach requires rigorous simulation and stress-testing of entire AI ecosystems, anticipating how autonomous agents might collectively respond to various stimuli, including those designed to be benign. The ethical imperative is to design for resilience and collective safety, not just individual agent performance, by embedding circuit breakers and human-in-the-loop oversight at the system level before widespread deployment.