When the Bank of England named AI-driven cyber risk as a top security concern for the banking sector, the response from much of the industry was familiar: patch more systems, and hunt for more vulnerabilities and weak spots. These are sensible measures. But they treat AI adoption as a technical problem to be managed, instead of the structural issue it is, says Dalit Amitai, Head of Product and Technology – Risk and Fraud Solutions, Bottomline.
We hear all the time that organisations are adopting technology faster than regulation can keep pace with. The rapid rollout of advanced models like Anthropicโs Mythos shows just how fast that capability curve is climbing. But while regulators and financial institutions debate how to keep up with legitimate AI use within the industry, who is looking at the criminals who are using AI within their own industry, and how does the step change to agentic AI play into this?
Criminals Do Not Have a Governance Problem
A bank rolling out a new AI tool understandably needs to get it past a compliance team, a risk committee and sometimes months of governance and policy reviews. Criminal groups are not constrained by these controls. That means they can pick up and test new tools almost as soon as they emerge, using AI to improve anonymity, scale and reach across phishing, money laundering, social engineering and increasingly sophisticated fraud.
The danger is that cybercrime and fraud are becoming industrialised, meaning faster, more convincing, and harder to trace attacks. For example, a fraud ring no longer needs a skilled voice actor to impersonate a customer over the phone โ AI voice clones can mimic voices well enough to fool verification checks, easily and at scale.
Bolting AI Onto Legacy Systems Creates the Gaps for Criminals to Exploit
Meanwhile, many financial institutions are responding to this pressure in the wrong way. Faced with the need to demonstrate AI adoption, the path of least resistance is to layer new capability onto legacy infrastructure. The problem is that bolting new tech onto old systems can add new weak points, more complexity and more compliance blind spots.
This is playing out on both sides. Criminals are using generative models to improve their systems of deepfake voice and video scams, increasingly convincing social engineering, and synthetic identity fraud. Banks, in turn, are deploying AI for their own transaction monitoring, identity verification and anomaly detection. But how effective those models are depends on the breadth of fraud signals available to them. Fraud moves across institutions, networks and borders, which means the defence has to as well.
Agentic and generative AI dramatically reduce the cost and expertise required to launch fraud attacks, allowing criminals to adapt their tactics in near real time. As fraud evolves more quickly, AI models trained on data from a single institution become outdated more quickly. Effective AI-driven fraud detection depends not only on better models, but also on access to broader, more current intelligence from across the financial ecosystem.
Why Patching Will Not Close the Gap
Patching systems and tightening governance are necessary, but they only address risk within the walls of one institution or firm. Fraud rarely respects those walls. The increasing speed of payments, particularly real-time and cross-border transactions, has exposed weaknesses in traditional financial controls. A single, coordinated scam can hit several banks, payment networks, and countries all at once. If each institution only sees its own picture, its fraud detection models are working with only a fraction of the available intelligence. That means it ends up reacting to patterns that another institution has already identified and possibly solved months earlier.
Collaboration Has to Become Normal, Not Exceptional
This is why collaboration and shared intelligence need to move from industry talking points to reality. Despite competing fiercely in every other respect, banks have a shared interest in mitigating fraud collectively: it lowers risk for the whole system and means fewer customers are wrongly flagged or blocked by overly cautious, siloed fraud checks.
Having said that, shared intelligence is only useful if it preserves customer privacy, complies with cross-border data regulation, and moves at a speed close to real time. Built correctly, a consortium approach to sharing intelligence gives AI-based fraud detection a broader, more current view of emerging threats. A collective defence helps institutions spot emerging attack patterns before they reach them, instead of picking up the pieces afterwards.
Collective Defence Is the Missing Link
Banks will continue adopting AI and will need to manage the risks that come with it. But patching vulnerabilities and tightening governance, institution by institution, will not address fraud that moves across the wider financial system. If the industry wants to keep up with fraudsters, collaboration and shared intelligence need to become an integral layer of its defence. Criminals are already operating as a connected ecosystem, so banks need to do the same.
