AI and financial stability
AI Systemic risk Policy
There are considerable concerns about AI making it easier to hack the financial system. AI helps attackers find vulnerabilities and carry out attacks more efficiently, as the UK's National Cyber Security Centre has assessed.
These concerns are justified. The wider financial stability effects depend on how banks use AI and how the authorities respond.
We are often told that a human will remain in the loop to approve decisions, or on the loop to monitor them and intervene when needed. As decisions become faster and more complicated, that person may increasingly depend on another AI to understand what is happening.
That AI may decide which problems reach the human and recommend what to do about them. The human may remain formally responsible while AI performs much of the effective oversight.
The banks' advantage over their supervisors matters, as does the way AI changes who can compete in banking. Both affect how banks behave in a crisis. For cyber attacks, the timing and the willingness of governments to cooperate may determine whether the damage becomes systemic.
Banks and supervisors
Banks have a direct financial incentive to adopt AI. A bank that can assess borrowers, manage liquidity or meet reporting requirements more cheaply has an advantage over its competitors.
The authorities face tighter budgets, procurement constraints and restrictions on how they use confidential information. They also lack the competitive pressure that makes delay expensive for a bank.
Banks can use AI to optimise against the rulebook. A bank might find a way to take on leverage that the regulatory measures fail to capture, or construct a portfolio that passes a stress test while retaining the underlying exposure. The reports improve while the risk stays put.
The supervisor may then be using a slow, human process to assess an account of the bank produced by a much more capable machine.
A bank's engine has access to its internal decisions and positions. The supervisor depends on what is reported and on its ability to interpret it.
If both buy from the same small group of suppliers, they may also share blind spots. The supervisor becomes part of the same risk monoculture it is supposed to oversee.
Winners and losers
AI also changes which banks can compete. As I discussed in my May 2025 SUERF piece, the largest banks may gain because they can afford to develop their own AI. But newer banks can have an advantage too, with modern systems and staff accustomed to using the technology.
The banks in the middle may struggle. They have older systems to maintain and less money to invest in AI. Buying from a vendor gives them access to the technology, but they still have to make it work with their systems and staff.
The result may be a more concentrated banking system, with the largest banks becoming even more important. Dependence on a small group of AI suppliers can also make banks' assessments of risk more alike.
Banks in a crisis
For 999 days out of a thousand, banks compete for profit. On the thousandth, survival takes over. They stop renewing loans, sell assets and conserve cash. Getting out early protects an institution. Getting out late can destroy it.
AI makes that race faster. As banks delegate decisions about liquidity, collateral and counterparty exposures, their systems can respond to signs of trouble before a human committee has met.
One bank withdraws funding. Others observe the withdrawal, infer that something is wrong and protect themselves. Their response makes the first bank's fears more likely to come true.
The engines need no agreement or direct communication. Prices and funding flows carry the information. Similar models, data and objectives make it more likely that many banks draw the same conclusion and act together.
Each bank can be managing its own risk competently while collectively draining liquidity from the system.
Andreas Uthemann and I have analysed how AI accelerates this process. A crisis that once unfolded over days may develop in hours, leaving less time for the authorities to organise a response.
Nor can the authorities assume their own AI will guide them through it. Successful performance in ordinary markets gives little assurance about decisions in an unprecedented crisis. AI can be least reliable when needed the most.
The double coincidence
The state of the financial system when a cyber attack arrives is crucial.
Suppose an attack disables part of the payment system on an otherwise calm day. Banks have liquidity, counterparties trust one another and the authorities can concentrate on restoring service. The incident may be expensive and disruptive, but the system has room to absorb it.
Let the same attack arrive during a dash for cash. A bank expecting a payment does not receive it. It cannot tell whether its counterparty has an IT problem or is about to fail, and may not have time to find out. It conserves its own cash, passing the shortage to the next institution. The operational failure feeds the panic.
This is the double coincidence, a cyber attack coinciding with financial vulnerability.
The authorities are already stretched, and an attack on payment or settlement infrastructure can obstruct the very channels through which emergency liquidity must reach the institutions that need it.
The record of attacks contained in normal times offers little reassurance about this combination. Restoring a service after a few hours may be a success on an ordinary day and far too late during a run.
Geopolitics and multilateral authorities
A hostile state with the necessary access and patience can wait for financial stress. It may also try to create that stress.
A criminal seeking payment usually wants a victim capable of paying. A state may want prolonged disruption and a government that appears unable to protect its financial system. Even restoring the damaged systems may not restore confidence if further attacks are expected.
Here I see a particular problem for the multilateral authorities. The IMF calls for international cooperation to address cyber risk. Cooperation becomes harder when a state expected to participate in the response may be sponsoring the attack.
The institutions assembled to protect financial stability can include governments that see advantage in another member's instability.
The Financial Stability Board relies on persuasion and peer pressure to secure implementation of its policies. Those tools depend on some common purpose. They have limited force against a government willing to damage another country's financial system for strategic reasons.
Identifying an attacker may depend on intelligence that governments will not share widely. Revealing a vulnerability may help an adversary. Uncertain attribution makes a joint diagnosis harder, while naming a suspected state can turn a technical response into a diplomatic confrontation.
National authorities can work with their security services and trusted partners. Multilateral bodies must work across the political divisions themselves.
My concern is that this pushes the multilateral discussion towards the risks on which members can agree, such as reporting standards and operational resilience, while making the deliberate use of financial instability harder to address. The resulting framework can look comprehensive even though cooperation may fail in a crisis.
Conclusion
The authorities should prepare for banks withdrawing liquidity while payment and settlement systems are disrupted. That includes making sure emergency liquidity can still reach the institutions that need it.
International plans must allow for governments withholding information or actively obstructing the response.