AI & Financial Stability – The Industrialisation of Cyber Risk

Yet again the news is awash with AI, Cyber Risk and Financial Stability FUD (fear, uncertainty doubt) but to cut through the verbiage, the attack path really has not changed.

Whilst it is fair to say regulators are right to be worried about AI enabled cyber risk in financial services, the concern needs to be framed carefully. The danger is not that modern AI models have suddenly invented an entirely new cyber universe. The more credible risk is that they make the existing one faster, cheaper and more scalable.

On one side of the argument, the European Systemic Risk Board and other regulators have a legitimate concern. Banks operate in a uniquely exposed environment. They are high-value targets, have complex supply chains, legacy systems, slow patch cycles, regulated change windows and deep dependency on common software providers. If attackers can use AI to automate reconnaissance, code analysis, exploit adaptation, phishing, vulnerability chaining and operational decision making, the balance of effort shifts. Weaknesses that once required a skilled team and weeks of work may become accessible to smaller groups in days or hours. That matters for financial stability because banks do not fail in isolation. Shared platforms, cloud services, identity systems, payment rails, managed service providers and open source components can create common mode weaknesses across the sector.

However, the opposing argument is also important. Most AI models are trained on the current body of knowledge. They do not magically create new categories of vulnerability. SQL injection, insecure APIs, weak identity controls, exposed management interfaces, poor segmentation, stale software, misconfigured cloud services and business logic flaws already exist. The core attack path remains familiar, such as identity compromise, privilege escalation, lateral movement, persistence, data theft, fraud or disruption. In that sense, AI does not fundamentally rewrite the attacker’s playbook. It industrialises it.

That distinction matters. If organisations respond as though AI has created a mysterious new threat class, they risk wasting effort on abstract controls, policy theatre and speculative scenarios. If they dismiss the issue because the vulnerabilities were already there, they miss the real point, AI changes attacker economics. The attack path may not be new but the velocity, volume and affordability of attack may be.

The logical mitigation approach is therefore not to create a separate AI cyber panic programme. It is to harden the existing cyber operating model against AI speed adversaries.

Banks should start by identifying the attack paths that would matter most at systemic level, internet facing systems, privileged access, identity federation, payment operations, third party connectivity, managed service access, software supply chain exposure and critical legacy platforms. They should then test whether vulnerability discovery, validation, prioritisation, patching and compensating control deployment can keep pace with accelerated attacker discovery.

The practical control agenda should include continuous external attack surface management, AI assisted secure code review, aggressive remediation of known exploited vulnerabilities, tighter privileged access controls, stronger segmentation, resilience testing of critical suppliers and pre-agreed emergency change processes. Just as importantly, banks need evidence of which risks were found, who accepted them, what was fixed, what remains exposed and why.

AI has not made every attack path new. It has made many old attack paths newly economic. That is why the regulatory concern is credible and why the answer is not more paperwork but faster assurance, faster remediation and a clearer view of where systemic exposure really sits. Faster and in real time, all the time and on demand. And that presents a challenge for rigid compliance regimes and traditional risk management, as I wrote about before.

So my giveaway For the banks is a practical question orientation. Do not ask yourselves “Do we have AI cyber risk on the risk register?” But, “Can we detect, prioritise, validate, patch and evidence remediation at the same speed that AI can discover and chain exploitable weaknesses?”

For many legacy heavy institutions, the honest answer is probably no and therein lies the simple truth. The attackers are operating in a real time AI world, while too many financial institutions remain trapped in a theatre of ticket queues, supplier dependencies, regulatory choreography and risk dashboards that turn red only after the fire has reached the boardroom. This is not technological fashion, it is the new minimum standard for operational resilience.