Banks now have minutes, not weeks, to patch — and always-on money is why that cost is structural

A warning out of Basel in early September sounds like the plot of a thriller: frontier AI has compressed the time a bank has to fix a software flaw from weeks to minutes, and routine patching schedules no longer keep pace. The source is the Bank for International Settlements — the "central bank of central banks" — through its Financial Stability Institute, in a paper on frontier AI threats to the financial sector. It identifies "autonomous vulnerability discovery and exploitation" as the single most significant development AI brings to banking. The terrorism of the headline is real, but it points at something slower and more durable for anyone holding bank stocks: cyber risk is quietly becoming an operating cost and a tail risk that most bank valuations barely price in.

Start with what the warning actually means in operational terms. A flaw is a bug in a bank's software that attackers can turn into access. For years the working assumption was that a bank had weeks — a monthly patch cycle, a scheduled maintenance window — between learning about a flaw and anyone weaponizing it. That assumption is what AI breaks. When a model can scan code, identify the weak point, build an exploit, and chain several weaknesses into one attack on its own, the attacker no longer needs the time or the specialized human skill the old model assumed. The BIS paper tells banks the discovery-to-exploit window has gone from weeks to minutes. Regulators and industry bodies are saying the same thing in their own numbers: the UK's Cross Market Operational Resilience Group expects repair timelines to shrink from weeks to days and, in some cases, hours, while the Institute of International Finance is urging banks to patch faster, outside scheduled windows, and to accept planned downtime for urgent fixes.

That last phrase is where the investment story hides. Accepting planned downtime sounds benign, but for a bank, downtime is when it is not doing the thing it is paid for — moving money, clearing checks, funding loans. Every year a bank spends more simultaneously defending against attackers and accepting the outages those defenses require. Look at the cost baseline. Financial services entered 2026 as the most attacked industry on the internet, and IBM's cost-of-breach work put the average financial-services breach at $5.56 million in 2025, second only to healthcare. Industry-wide, Gartner expected global cybersecurity spending to reach about $240 billion in 2026, up more than 12% from the prior year. And the asymmetry on display in this report is what makes the trend sticky: as BIS itself has argued, AI lowers the cost of sophisticated attacks more than it lowers the cost of defense, so attackers get cheaper while defenders get busier.

Here is the part the headline does not tell you, and it is the part worth thinking about structurally. The very trend that makes "minutes not weeks" urgent is the one pushing banks toward always-on money — and always-on money cannot be paused for a patch. Throughout the same period, banks have been racing to modernize onto real-time rails: instant payments, tokenized deposits, and 24/7 settlement. BIS itself is a driver of this. Its Project Agorá — a collaboration of seven central banks and more than forty financial institutions — is testing tokenized wholesale settlement, and in 2026 that moved beyond prototype into real-money tests, with TD moving actual U.S. dollars and a live cross-border pilot involving JPMorgan, Citi, and UBS. All of it built around one premise the BIS warning quietly undermines: that settlement could run continuously, in near-real time, forever.

A monthly patch cycle is a compromise a batch-processing world could afford, because a system that runs at scheduled intervals can also be taken down at scheduled intervals. A real-time settlement rail that you have promised customers is available 24/7 is a system you cannot simply switch off every time a new vulnerability appears. Either you keep it up and carry the risk, or you take it down and break the promise that made it valuable. That is a genuine tension, not a rhetorical one, and it means the pressure this report describes does not stay where the report found it. It compounds as money moves further onto always-on infrastructure. The institution warning about un-patchable speed is, at the same time, building the infrastructure that assumes downtime is no longer an option.

For a retail investor, the honest read is not "sell banks because someone might hack them." No single breach is predicted here. The read is a repricing of two things a bank's price-to-book does not show you: a slowly rising operating cost that lands on every margin, and a tail risk that becomes more probable as the fence gets harder to hold. Both are structural, which means they reward the banks best placed to absorb them. A bank running a modern, cloud-native, always-on stack can patch continuously and treat downtime as part of design; a bank still running a decades-old core processor, stitched together with bolted-on security, faces urgent patching that means real outages, real customer complaints, and real attrition of deposits and fee revenue at the exact moment the sector's margins are already under pressure.

The most useful question this warning raises is not whether your bank will be breached, but which kind of bank it is. The ones built to patch without pausing have turned a cost into something closer to a moat. The ones that cannot are carrying a risk that does not appear on Wednesday's statement. As money moves onto rails that never sleep, that distinction is only going to get more expensive to ignore.