A run, at machine speed.
By Mara Quinn | November 17, 2026
A Warning From Inside the System
The Federal Reserve does not often flag speculative threats. It monitors what it can measure: credit spreads, reserve ratios, capital buffers. When a Fed official stands up and names something as a potential trigger for bank runs, that is worth paying attention to.
That is what happened this week. Fed official Michael Barr warned publicly that AI agents, software programs designed to act autonomously on financial instructions, could fuel future bank runs. The warning appeared in multiple outlets on November 16, 2026, including Bloomberg and Gizmodo. By November 17, the South China Morning Post had added its own reporting on the broader risk picture around agentic AI in finance.
Barr's concern is not that AI will make bad investment calls. Regulators already understand that risk and know how to frame it. His concern is structural: the speed and coordination of AI agents, acting together on the same signal at the same moment, could overwhelm the systems banks rely on to calm depositor panic before it compounds.
The Mechanism: Speed Outpacing Reassurance
Bank runs have always been about speed. The classic panic unfolds in a recognizable sequence: news breaks, depositors worry, some withdraw funds, others see the withdrawals and withdraw too, and the spiral tightens faster than the institution can respond. What traditionally slowed that spiral was friction. Humans check news, call their bank, wait for an ATM, reconsider. The lag gave banks time to issue statements, regulators time to intervene, and calm depositors time to steady nervous ones.
AI agents remove that friction almost entirely.
An agent managing a customer's finances can be programmed to monitor news feeds, evaluate sentiment, and move funds if certain conditions appear. A thousand such agents, trained on similar logic, watching the same data sources, can reach the same conclusion at roughly the same moment. They do not need to call a hotline. They do not sleep on it. They act.
Barr and others at the Fed appear to fear a scenario where many agents, acting simultaneously on a shared signal, drain deposits from an institution faster than any human communication effort can counteract. A bank spokesperson issuing a reassurance statement operates on a timescale of hours. An AI agent operates on a timescale of milliseconds. The asymmetry is total.
That is a direct consequence of how agentic systems are designed: to execute quickly, follow instructions precisely, and not second-guess the logic they were given.
Signal Propagation and the Echo Chamber Problem
The speed problem is compounded by how AI agents gather information. Many pull from overlapping data sources: financial news aggregators, market feeds, social media. If a false or misleading signal enters one of those feeds, it propagates to every agent subscribed to it. The agents do not fact-check. They act.
Traditional depositor behavior includes individual variation. Some people read multiple sources. Some call their branch manager. Some decide to wait. That variation creates natural damping in the system. When individual depositors are replaced by software following similar rules, the variation shrinks. The system grows more brittle, not more stable.
The South China Morning Post's coverage on November 17 pointed to this dynamic in its examination of blind spots in agentic AI applied to finance. The concern is not just speed, but the reduction in behavioral diversity that agentic systems may introduce. A market of thousands of human depositors making independent decisions is relatively hard to tip into synchronized panic. A market of AI agents optimizing similar objectives on similar data is considerably easier.
Beyond Bank Runs: The Rogue Agent Problem
Barr's warning focused on coordinated action by agents doing exactly what they were programmed to do. A separate concern, raised by BankInfoSecurity on November 16, involves agents doing something their designers did not intend.
Security researchers and financial institution analysts have begun pushing for formalized threat modeling specifically around AI agents operating in financial environments. The concern is that rogue agents, whether compromised, manipulated, or simply behaving outside their tested parameters, represent a category of risk that existing security frameworks are not built to handle.
Traditional cybersecurity threat modeling asks: who might attack, what are their tools, what are they after. Rogue AI agent threat modeling adds a harder question: what happens when the source of harm is software your own institution deployed, acting in ways no human directly authorized.
The BankInfoSecurity reporting framed this as a gap in current security practice. Threat models for AI agents in finance are, in many institutions, either absent or treated as an extension of existing software vulnerability frameworks. That approach misses the specific dynamics of autonomous action: agents that interact with external data, make decisions, and execute transactions without pausing for human review at each step.
The Insider Threat Analogy
Security analysts have drawn a rough analogy to insider threats. An insider threat is dangerous not because it comes from outside the perimeter, but because it operates with legitimate credentials, in expected systems, doing things that look like normal activity until they do not. A rogue AI agent fits a similar profile. It has access. It has authority. Its activity looks like intended behavior. The difference between a correctly functioning agent and a compromised one may not surface in logs until significant damage has already occurred.
That analogy has limits. Insider threats involve human intent. AI agents do not have intent in any meaningful sense. But the structural similarity is useful: both categories exploit legitimate access, both are hard to detect through perimeter defenses, and both require a different kind of monitoring than external attacks demand.
What Regulators Are Watching
Barr's public statement appears to be part of a broader Fed effort to get ahead of the agentic AI curve in financial services. The warnings are coming now, before widespread deployment, which suggests a degree of intentionality in the communication. This is not a post-crisis assessment. It is a prospective one.
The Defiant's reporting on Barr's remarks, also from November 16, framed the concern in terms of acceleration. Barr did not say AI agents would cause bank runs in the way a fraudulent bank statement might. He said they could accelerate the dynamics that cause bank runs, compressing timelines that regulators and institutions have historically had room to manage.
That framing matters. The Fed is not suggesting AI agents are bad actors. It is suggesting that agents operating as intended, in a world where many institutions deploy them simultaneously, could create emergent behavior that no individual institution or agent designer anticipated. The system becomes the problem. The parts work correctly. The whole fails.
Gizmodo's coverage described the Fed's concern in similar terms: not rogue behavior, but coordinated correct behavior at a speed the existing financial infrastructure cannot absorb.
The Gap Between Deployment and Oversight
One pattern in the reporting from this week is the distance between how quickly financial institutions are deploying agentic AI and how slowly the regulatory and oversight infrastructure is developing to match. The South China Morning Post's reporting pointed to new risks and blind spots that existing frameworks do not cover. Those blind spots are not primarily technical. They are procedural and organizational.
A financial institution that deploys AI agents for account management, trading, or customer service may have excellent technical documentation for each system. What it may lack is a framework for thinking about how those systems interact with each other, with external agents at other institutions, and with market conditions that none of the individual systems was designed to anticipate. The aggregate behavior of well-designed individual agents is not automatically well-designed behavior at the system level.
Barr's warning is significant. The Fed is pointing at a systemic vulnerability. The risk does not live in any single agent or any single bank. It lives in the interaction between agents, institutions, and the signals they all process at the same time.
What the Industry Needs to Do
None of the reporting from this week describes concrete regulatory proposals. Barr's warning appears to be a public statement of concern rather than a preview of specific rule changes. That leaves a lot of open space.
The BankInfoSecurity reporting suggests that threat modeling is the minimum starting point. Financial institutions deploying AI agents need frameworks that explicitly address autonomous action, the propagation of bad signals through agentic networks, and the detection of behavior that deviates from intended parameters. Building those frameworks while deployment continues is harder than building them before deployment scales. The industry may already be past the easier window.
The other gap is at the infrastructure level. Banks have communication systems designed for human timescales. If an AI-driven withdrawal event compressed what would normally be a 24-hour crisis into a 20-minute one, existing communication infrastructure, the press statements, the regulator calls, the FDIC coordination, would not keep pace. Some of that infrastructure may need to be rebuilt around the assumption that the next candidate for a bank run will move much faster than the last one.
Agentic AI in financial services is not going away. The efficiency arguments for deploying it are real: automated account management, faster transaction processing, continuous fraud monitoring. This week's reporting makes clear that the same properties producing those benefits - speed, autonomy, coordination across shared data - also produce a category of systemic risk the financial system is not currently equipped to handle.
Barr said it out loud. The question now is whether the institutions and regulators listening will act before the question becomes empirical.
Sources
- Bloomberg — Fed's Barr Says AI Agents Could Fuel Future Bank Runs, November 16, 2026
- Gizmodo — The Fed Is Worried AI Agents Could Trigger a Bank Run, November 16, 2026
- The Defiant — Fed's Barr Warns AI Agents Could Accelerate Bank Runs, November 16, 2026
- South China Morning Post — Agentic AI in finance: new risks and blind spots, November 2026
- BankInfoSecurity — Rogue AI Agents Make Threat Modeling a Security Must-Have, November 16, 2026

