The Fed's new stability report says the next crisis may start in the payment pipes, not the stock market. What breaks first when the messages stop.
7 min read
THE SUNDAY EDITION
The Federal Reserve is worried about artificial intelligence. The most interesting risk may be hiding in the financial plumbing.
There is a peculiar comfort in a stock market crash. Not while it’s happening, obviously. But at least we understand the mechanics. Someone paid too much for something, reality eventually intervened, and a great deal of money disappeared from screens.
Painful. Occasionally spectacular. But familiar.
A stranger financial crisis would begin with everyone still having their money, at least on paper, and suddenly discovering they couldn’t move it.
That possibility deserves attention after Friday’s Federal Reserve Financial Stability Report. In a survey of 23 market contacts, 87% cited artificial intelligence as a potential shock to financial stability over the next 12 to 18 months, up from 50% in the spring survey. Those were the respondents’ concerns, not an official Fed forecast. And their concerns included stretched valuations and complicated AI financing, not just cybersecurity. [1]
Wall Street has spent months debating whether the AI boom is worth its price tag. Are the revenues real? Will the data centers pay for themselves? How many chips does civilization need before someone asks for a receipt?
Fair questions. But there is another one: what if AI creates trouble even when the technology works commercially?
The financial system has two very different jobs. One is deciding what things are worth. The other is making sure money gets where it is supposed to go. We spend almost all our time talking about the first. The second is what makes the first possible.
Consider two pieces of infrastructure most Americans have never heard of: Swift and Fedwire.
Swift is the global network banks use to exchange standardized financial messages, including payment instructions. It does not generally settle the money itself. Fedwire Funds Service, operated by the Federal Reserve Banks, does something different: it settles large-value transfers between participating institutions in central bank money. [2][3]
The scale is almost comical. In 2025, Fedwire processed more than 217 million transfers worth approximately $1.148 quadrillion, averaging $4.59 trillion per business day. That is gross payment activity, not $1.148 quadrillion of newly created wealth. The same dollars can move repeatedly through the system. But the numbers tell you how much activity depends on its reliable operation. [3]
Swift is similarly hard to visualize. Its published overview describes more than 11,500 connected members and says its network communicates value equivalent to the world’s annual GDP roughly every three days. It reported more than 53 million payment messages per day in 2024. Again, these are messages and the value associated with them, not money held by Swift. [2]
These networks are supposed to be boring. In fact, boring is a design requirement.
Imagine a major bank discovers that an attacker has compromised the systems it uses to prepare or authorize payment instructions. It may need to suspend certain payments while it investigates. Transactions pile up. Counterparties wait for funds they expected to receive. A firm that counted on an incoming payment now has to find liquidity elsewhere.
The money has not necessarily vanished. The plumbing has stopped working.
This is a hypothetical scenario, not a claim that either network has been breached. The vulnerability may sit at the edge: in a participant’s credentials, internal payment applications, third-party technology, or the process that decides whether an instruction is legitimate. A resilient central network cannot automatically make every connected institution resilient.
That distinction is important. Swift has a customer security program and control framework, while Fedwire has formal operating and contingency arrangements. The question is not whether either organization forgot to buy antivirus software. It is whether the entire ecosystem can keep pace as attacks become faster, cheaper, and more automated. [4][5]
The Fed noted increasing reports of agentic AI systems being involved in cyber incidents since its previous report. That does not mean an autonomous model is about to empty every checking account. It means attackers may be able to automate more of the work involved in finding weaknesses and exploiting them. Defenders can use AI too. The arms race runs in both directions. [1]
Finance is a particularly unforgiving place to hold that race. A retailer can lose a day of sales during an outage. A bank’s delayed payment can become another institution’s funding problem. A clearing or settlement obligation does not politely disappear because somebody is waiting for the help desk.
And there is a second, quieter issue: concentration.
Suppose thousands of financial institutions independently choose the same handful of cloud providers, model vendors, or specialized software services. Each choice might be perfectly rational. Faster processing, lower costs, fewer people reconciling spreadsheets on a Friday evening. Multiply those decisions across an industry, however, and individual efficiency can become collective dependence.
A thousand banks connected to resilient payment networks still represent a thousand separate points of operational vulnerability. And if many of them share a critical technology supplier, they may also share the same potential point of failure.
This is not an argument against AI. It is an argument against confusing convenience with resilience.
The financial industry already invests heavily in cybersecurity, continuity planning, and recovery capabilities. New AI tools could make those defenses better. Nothing in the Fed’s report establishes that an AI-driven payment crisis is imminent, or that Swift or Fedwire is uniquely vulnerable. The more useful lesson is about the standard we apply to systems once they become indispensable.
A chatbot can hallucinate a restaurant recommendation and apologize. A model helping process a critical payment instruction has considerably less room for improvisation.
Technology often follows the same path. First it is an interesting tool. Then it becomes part of the workflow. Eventually everyone forgets how to work without it. The dependency becomes visible only when something breaks.
That is how convenience becomes infrastructure.
The headline number in Friday’s survey was 87%. It captured the financial industry’s growing unease with AI, its valuations, its financing, and its risks. But the more interesting question may be what happens after everyone starts relying on it.
For years, financial technology has promised to move money faster, more cheaply, and with less human intervention. We are getting better at all three.
Now comes the harder test: making the system more intelligent without making it more fragile.
Because the most expensive thing about financial infrastructure is not always what it costs to build. Sometimes it is discovering how much everyone depends on it.
— Greg Ranieri, Editor-in-Chief
References and fact-check notes
1. Federal Reserve Board, Financial Stability Report (November 13, 2026), especially Box 5.1 and discussion of cyber risk.
https://www.federalreserve.gov/publications/files/financial-stability-report-20261113.pdf
2. Swift, “What is Swift?” (network function, scale and messaging statistics; the 53 million figure is for 2024).
https://www.swift.com/about-us/who-we-are/what-swift
3. Federal Reserve Financial Services, “Fedwire Funds Service – Annual Statistics” (2025 annual value, volume and daily averages).
https://www.frbservices.org/resources/financial-services/wires/volume-value-stats/fedwire-funds-service-annual-statistics/
4. Swift, Customer Security Programme (security expectations for connected institutions).
https://www.swift.com/myswift/customer-security-programme
5. Federal Reserve Financial Services, Fedwire Funds Service (service and operational documentation).
https://www.frbservices.org/financial-services/fedwire-funds-service/
Editorial verification: The 87% result reflects a 23-person survey, not a Fed probability estimate. Swift message values are not funds settled by Swift. Fedwire values represent gross transfers. The cyberattack and payment-interruption examples are illustrative, not reported incidents.
KEEP READING
The Planes Stopped. The Bill Kept Growing.
Inflation Said Hike, Nvidia Lost Its Chief, the Market Rose Anyway
New posts by email
One short email when a new Financial Spark story publishes. No newsletter, no spam — unsubscribe anytime.



