Unravelling the new economics of attacking a bank
Agents do not need to invent an entirely new form of bank fraud. They only need to make existing methods of attack faster and cheaper.
Picture a bank as a castle. Over time, the largest banking castles have been extended. New wings have been added, old rooms repurposed and passages constructed to connect one generation of technology with another. Customers enter these banking castles through mobile apps and websites, while employees, suppliers, payment companies and other banks use a multitude of different entrances. There have been tons of structural changes over the years, meaning some older doors are no longer in use, but they are not blocked-up because an essential service still depends on them.
The walls around these banking castles are formidable because banks have spent a lot of time and money on defending themselves. They employ sophisticated security teams and use behavioural analytics, device intelligence, and automated monitoring to protect customers. They've been fighting bots for years and they are really good at it too. The issue is, the game has now changed substantially and that is down to AI agents.
Agents do not need to invent an entirely new form of bank fraud. To succeed, all they need to do is to make existing methods of attack faster and cheaper. The National Cyber Security Centre has said it expects AI's greatest near-term effect to be the acceleration and scaling of familiar attack techniques, rather than the sudden appearance of wholly new ones.
Take the conventional bot, something we have all been familiar with for some time. Bots are usually given a predetermined task. One might test thousands of stolen passwords, submit fraudulent account applications, or try small payments against compromised cards. When one approach is blocked it stops, repeats the same action, or waits for its operator to change the instructions. A capable agent, though, can potentially behave differently. It can observe the response, adjust its approach, and try another route. If it cannot get through the login page, it might examine the password reset process. If that fails, it could try the mobile interface, interact with customer support, or search for exposed information about the account holder. It could study the bank's error messages, shift its behaviour, and return later with a better story.
An attack that once required a team of bad actors to orchestrate it could increasingly be conducted by one person directing multiple agents. It becomes ever more worthwhile if the labour cost is almost nothing. For criminals, a very low success rate can still produce a return if millions of attempts can be made cheaply enough.
Every door is an opportunity
The UK's financial authorities have already warned that AI could identify and exploit large numbers of vulnerabilities more quickly. They are concerned that there's not an easy fix for organisations that provide access to money, personal information, and valuable commercial data. The difficulty is that banks must remain continuously available, connect with large numbers of outside organisations and be able to distinguish fraudulent activity from legitimate customers who simply behave in unusual ways. There is a particular threat to the largest established institutions with their substantial and complicated technology estates. Bad actors don't need to find a weakness in the main banking app. They could exist in an older customer journey, a supplier connection, an account recovery process, or an internal administrator system. In other words, a criminal does not need to knock down the strongest wall in the castle. An agent can continuously search for an overlooked window in an extension built 20 years ago. Indeed, thousands of agents can test variations of the same approach simultaneously. When one route closes, they can look for another. This is why the Bank of England has been exploring the idea of a 'minimum viable bank', a protected, stripped-back set of functions, capable of keeping essential services running during a serious cyberattack. It is sensible resilience planning, but also an acknowledgement that a large bank cannot assume its entire technology estate will remain available during an attack. In castle terms, this is the equivalent to the incumbents building a defensible keep within the existing walls.
Established banks are not defenceless. They have large security budgets, experienced teams, extensive intelligence about previous attacks and years of customer behaviour data. The challenge lies in defending every part of an estate that has grown over decades. Essential software is difficult to patch without interrupting customer services. Firms also need to find, prioritise and remedy weaknesses at a pace that's difficult to sustain through human effort alone.
One closely guarded window
A bank built today has an opportunity to tackle this threat in a different way. It would still need apps, websites, APIs, payment connections and links to suppliers, but only one 'window' is required, where every agent interaction passes through one consistent control layer.
The process would be something like this: an agent approaching the window would have to establish its identity, the provider behind it, the customer it represents, and the authority it has been given. Its request would be checked against a machine-readable mandate before it could reach the bank's critical systems. A shopping agent might, say, be permitted to spend £20 a week with approved supermarkets. An authorisation could be made by a small business agent to pay known suppliers up to an agreed amount. Neither action would automatically allow a new payee to be added or to change its own limit or gain access to an unrelated part of the account.
In this scenario, an agentic-focussed current account would not need to ask whether the request looked like something the customer might do. It could establish whether that particular agent was authorised to do it. That is important for security but also for convenience. A stolen agent credential could be disabled without cancelling the customer card, closing the account, or removing the authority of every other trusted agent. A request outside the mandate could be refused or escalated before any money moved out of the account. Meanwhile, every approach to the window would also leave a record of which agent made the request, what it asked to do, whose authority it used, and why the bank allowed, rejected, or challenged it.
Defensive agents that never stop watching
A current account built by agents shares the same advantage as the attackers; it works 24-7 at machine speed. Development agents can continuously examine code and generate adversarial tests. They can test authentication journeys and look for unintended ways of acquiring greater access. Agents can help engineers identify vulnerabilities before software is released and prioritise those that create the greatest risks.
This does not mean allowing defensive agents to make unlimited decisions. A security agent can also be wrong. It could block a legitimate customer, misinterpret unusual behaviour, or disrupt an important service. Defenders, therefore, need mandates of their own. They should have clearly defined authority to observe, slow, isolate, or refuse. More consequential actions should still be escalated to skilled people.
Building an agentic current account from scratch represents an architectural opportunity. When you start with fewer entrances, or even just one, it is easier to defend against fraud. It can begin with a smaller technology estate, consistent identity controls, new narrowly scoped credentials and a clear separation between systems. Both legitimate customer agents and hostile agents can be treated as founding assumptions rather than additions to be patched later. This is the thinking behind a bank built by agents, for your agents and defended by agents.
