The next banking crisis may start outside a bank
The next banking crisis may not begin with bad loans, a run on deposits, or a failing bank. It may begin with a cyberattack or a technology failure outside the regulated balance sheet.
That warning came from Reserve Bank of India Governor Sanjay Malhotra in an October 3 speech on financial stability. The speech is not a new regulation, but it is an important policy signal: AI, cyber risk, third-party dependence, market valuations, and financial contagion can no longer be tested as unrelated problems.
Most institutions still divide them among separate owners and separate exercises:
- security tests compromise and recovery;
- operational resilience tests service continuity;
- model risk tests AI behaviour;
- market risk tests price shocks;
- treasury tests liquidity;
- third-party risk reviews vendors.
The organization chart is tidy. The real failure path is not.
One event can cross every risk boundary
Consider a plausible scenario.
A widely used technology provider is compromised. Several banks lose access to a critical service. Payment and customer-service capacity falls. Uncertainty spreads because nobody initially knows which institutions, data, or dependencies are affected.
Customers move funds. Corporate clients draw committed facilities. Markets reprice exposed technology companies and financial institutions. Leveraged investors face margin calls and sell liquid assets. Collateral values fall. Banks and non-bank financial institutions demand liquidity at the same time.
No single step needs to destroy the system. The danger comes from amplification across operational, market, funding, and confidence channels.
This is the central architectural implication of the RBI speech. “Cyber Risks Compounded by AI” was one of five emerging global financial-stability risks identified by Malhotra, alongside stretched asset valuations—particularly AI-related valuations—elevated leverage, global debt, and private credit. He warned that these shocks occurring simultaneously could put significant pressure on the global financial architecture.
AI appears in this model twice.
First, it can increase cyber risk, model risk, third-party dependence, and the erosion of human oversight and accountability. Second, the AI investment cycle has become a market-risk factor: weaker investment or earnings could trigger a repricing of AI-linked assets, with leverage amplifying the correction.
Treating those as independent “technology” and “market” issues misses the possibility that one event activates both.
Dependency maps must describe consequences, not vendors
A conventional third-party register usually answers administrative questions: who the vendor is, what contract exists, when the assessment expires, and which service owner is accountable.
A systemic dependency map needs to answer harder questions:
- Which customer journeys, payment flows, risk decisions, and recovery processes depend on the provider?
- Which other providers, networks, cloud regions, identity systems, and data sources sit behind it?
- How many institutions depend on the same service or substitute?
- Can the service be replaced under stress, and how long would that actually take?
- What happens to payments, collateral, liquidity, and customer behaviour while it is unavailable?
- Which containment action can isolate the failure without disabling the entire institution?
The object being mapped is not a supplier. It is a chain from technical dependency to financial consequence.
That chain often includes entities outside direct banking supervision: cloud and identity providers, market-data services, payment processors, AI platforms, telecommunications networks, outsourced operations, and software supply chains. A bank can be individually well capitalized and still depend on infrastructure whose failure affects many institutions at once. Partner and outsourced channels are already a proven leak path: in Korea, customer data left banks through non-core systems.
Substitutability is therefore not a checkbox. If every bank plans to fail over to the same alternative provider, the industry has a common recovery assumption rather than genuine resilience.
Stress tests should connect four clocks
A useful cyber-originated systemic scenario must reconcile four different time horizons.
1. The technical clock
How quickly can the institution detect compromise, revoke access, isolate systems, restore trusted data, and recover service? Recovery time measured from incident declaration is misleading if detection took days.
2. The customer clock
How quickly do failed payments, unavailable balances, delayed support, or public uncertainty change customer behaviour? Confidence can move faster than infrastructure recovery.
3. The market clock
When do asset prices, collateral haircuts, margin requirements, and counterparty limits react? Automated trading and risk systems can propagate a change before the operational incident is understood.
4. The liquidity clock
When do deposit outflows, facility drawdowns, settlement obligations, collateral calls, and intraday liquidity needs peak? A system restored tomorrow may not solve a liquidity requirement due today.
Testing each clock separately can produce four successful exercises and one failed institution. The scenario has to run them together.
AI resilience requires evidence, not a policy statement
The RBI framework also exposes a weakness in many AI-governance programs. A policy may define human oversight, model approval, and acceptable use while saying little about the evidence needed during a combined operational and financial event.
The evidence pack should show:
- which models and agents affected the decision or workflow;
- which identity authorized each action;
- which data and tools were available to the system;
- which policy version was enforced;
- where human approval was required and whether it occurred;
- how the system was contained, degraded, or disabled;
- which downstream decisions must be replayed or reconciled.
Without this evidence, “human oversight” is difficult to demonstrate after the fact, and recovery teams cannot reliably distinguish trusted outputs from decisions made during compromise or model failure.
AI governance, cyber response, and financial reconciliation therefore need a shared event record. Three disconnected audit trails are not enough.
The board-level question changes
The old resilience question was: can this bank survive a severe financial shock?
The newer question is broader: can the financial system continue delivering essential services when a non-bank technology failure creates operational disruption, market repricing, confidence loss, and liquidity demand at the same time?
That requires system-wide exercises spanning banks, non-bank financial institutions, payment systems, markets, critical technology providers, and cross-border networks. It also requires better data about shared dependencies and interconnected exposures—the exact gaps highlighted in the RBI speech.
The practical next step is not another standalone AI-risk workshop. It is one joined scenario with named providers, quantified dependencies, realistic recovery constraints, market and liquidity transmission paths, decision owners, and tested containment options.
If the crisis can cross organizational boundaries, the stress test must cross them too.