Financial sector resilience in the age of AI: regulatory trends and supervisory expectations

Digital operational resilience is becoming an increasingly important issue for the financial sector and its supervisors. This is due not only to ongoing technological developments, including in artificial intelligence, but also to recent cyber incidents affecting or targeting the sector. 

Supervisors were especially alarmed by the release of Mythos, Anthropic's frontier AI model, and its ability to quickly identify severe cyber vulnerabilities undetected by humans. This prompted governmental engagement with the US banking sector on their cyber resilience, and tighter controls on foreign access to Anthropic's AI models. 

In Europe, the focus on cyber resilience is also increasing. On 7 July 2026, the European Central Bank askeddemanded Eurozone banks to draw up plans to counter AI-enabled cyber threats. On the same day, the European Systemic Risk Board published a warning about systemic cyber risks stemming from frontier AI models. It stated that rapid AI developments are materially changing the risk landscape, which could undermine the functioning and trustworthiness of the financial system. 

Setting the scene: Lagarde's address to practice and what it means for institutions

More recently, on 1 October 2026, Christine Lagarde opened the annual ESRB conference with a speech in which she highlighted the risks that AI can mitigate but also generate. She drew attention to operational consequences of reduced human oversight and control, as well as increased threats and the ongoing weaknesses of financial institutions in harnessing technology to defend against cyber-attacks. These developments impact operational resilience, which the European Banking Authority defines as the ability of an institution to deliver critical operations through disruption. For financial institutions, operational resilience must build on their prudential operational risk framework, which encompasses internal governance, outsourcing, business continuity and relevant risk management aspects.

With so many publications, communications and developments emerging at a rapid pace, it can be difficult to keep up to date with what is required and expected, and to understand the margin of discretion available to institutions. In this article, we look at existing frameworks, supervisory expectations and key developments in the areas of operational and cyber resilience, as well as AI risk management. 

Ensuring business continuity and viability within existing frameworks

It is clear that institutions must manage cyber and AI risk not just as a matter of operational risk management but as a key component in ensuring their business continuity and viability. A serious attack could cause significant financial losses, as well as damage to reputation and customer trust, which could threaten the institution's licence to operate. However, the "hard" legal requirements are limited and somewhat open-ended. 

For institutions, improving operational resilience means having ''robust and regularly tested crisis management, incident response, backup, failover, restoration and recovery arrangements aligned with DORA requirements and best practices for ICT risk management, incident handling and operational resilience testing […] in a context of higher probability of ICT-related incidents materialising'', according to the ECB. As this statement illustrates, AI-related operational risks are currently managed within the parameters of existing frameworks. 

The primary source of targeted rules for financial institutions on operational resilience is the EU legislation on digital operational resilience for the financial sector (DORA). DORA obliges institutions to put in place a comprehensive ICT business continuity policy, setting out their response to ICT-related incidents and the estimated impact, as well as their communication and crisis management policies.

In addition to DORA, there is additional EU and national legislation to consider, especially the Cyber Resilience Act, the NIS2 Cybersecurity Directive, which has been implemented in the Netherlands through the Cybersecurity Act (Cbw), and the Critical Entities Resilience (CER) Directive, implemented in the Dutch Critical Entities Resilience Act (Wwke). However, financial institutions are largely exempt from these frameworks based on DORA applying as a lex specialis for the sector. Nevertheless, the focus on resilience in both pieces of legislation makes them relevant for financial institutions. 

In the Netherlands, designations of banks and financial market infrastructures as critical entities in accordance with the Critical Entities Resilience Act took place in September 2026. The impact of the designation is rather limited as financial institutions are excluded in light of DORA. However, there are certain obligations for the appointing authority - for the banking sector and financial market infrastructure, the Minister of Finance - including an obligation to support and cooperate with these critical entities to enhance their resilience, for example by developing guidelines, assist in testing their resilience and providing advice and training to their staff.

Financial institutions should not view resilience merely as a matter of legislative compliance, but as an integral part of their business model and licence to operate. This is illustrated by the revised SREP guidelines. As part of the periodic supervisory review and evaluation process, ICT risk is elaborated upon in much more detail under the operational risk category, which forms part of the assessment. Specifically, authorities should consider not only the institution's exposure to ICT risk, but also the impact on its critical or important functions, including from a financial, reputational, regulatory and strategic perspective. 

Once again, we see integration through existing SREP elements rather than the introduction of a separate, independent risk category. In the context of the SREP, this promotes consistency and simplification. The same is true of other resilience-related legislation, as the responsibilities and obligations of institutions and supervisors alike are embedded in existing frameworks. Individual institutions should also build upon frameworks, systems and processes already in place, in line with the ECB's call for a "'comprehensive action plan" to enhance cyber resilience. 

While addressing risks, institutions must also incorporate new technologies and AI into their business models. Technology is evolving rapidly and, as Patrick Montagner, member of the ECB Supervisory Board, has emphasized, institutions must keep up with innovation. They must adopt new technologies to remain competitive and robust. The ECB especially identifies the use of AI as a way to keep up with technological developments, describing how banks use AI to accelerate IT operations and legal and document analysis. 

One challenge is that, while AI can help institutions to identify weaknesses faster, potential fixes must be tested before being deployed in essential services. This slows down institutions that want to use AI to protect themselves against attacks. The Financial Stability Board (FSB) has noted that agentic AI is increasingly being adopted for the use of planning, explaining and even executing multiple tasks independently. Like the ECB, the FSB stresses the importance of adopting AI responsibly. However, this can be challenging for institutions due to the lack of a clear framework. 

Governance

The ECB has been explicit about the need for institutions to adapt and strengthen their governance and control systems in response to the changes in the cybersecurity landscape. In her recent "Dear CEO" letter, Claudia Buch, chair of the ECB Supervisory Board, stated that ''[r]esponsibility for responding to the evolving cyber-risk environment primarily lies with banks’ management bodies." Boards are expected to revisit strategic ICT-related decisions, including ICT investments, resource allocation and risk tolerance frameworks (RTFs).

But what does strengthening governance and control systems for cyber resilience actually involve? Supervisory authorities draw attention to different aspects:

• Education: The Bank of England highlights the importance of enhancing the understanding of frontier AI risks among boards and senior management, so that they can oversee how control functions manage these risks.
• Validation: The OECD has monitored how different supervisory authorities guide institutions on their AI governance, concluding that robust model validation, continuous monitoring and independent validation are of great importance where AI models influence business functions. The OECD also concluded that, for robust governance, it is necessary not only to save the latest version of a model, but also to retain each instance in order to create a record for traceability. 
• Explainability: EIOPA addresses the increasing need for transparency and explainability in the context of AI. it mentions how new data sources and data types in combination with the system's black box result in outcomes that may be difficult to explain or interpret. One way to achieve explainability is to add a higher level of human oversight through human-in-the-loop decision-making processes. EIOPA also advice developing meaningful and easy-to-understand explanations and transparent communications about the data used in AI models, as well as clear communication to consumers about whether or not the decisions are made using an AI system. Similarly, the Central Bank of Ireland has stated that AI systems without adequate human oversight may result in decisions that are "neither fair nor explainable to consumers". 
• Visibility and evaluation: The FSB emphasises the importance of maintaining visibility of the risks associated with AI adoption throughout the AI lifecycle. Specifically, the FSB mentions organisational adaptability, which can be improved by providing foundational AI training and continuous learning to ensure governance, controls and oversight are updated where needed. This means that monitoring must be ongoing and regularly adapted to ensure compatibility with evolving AI approaches and risk landscapes.
• Third-party providers: According to the FSB, a part of maintaining visibility of AI risks is to manage third-party AI providers. Especially when third parties manage model parameters and version updates, it is recommended that financial institutions establish contractual arrangements with the vendor to ensure the timely provision of information. The importance of establishing clear contractual arrangements with third parties is also relevant in the context of the concentration risks discussed below. 

In this respect, DORA obliges institutions to have an effective and prudent governance and control framework. This must set out the management body's responsibilities for bearing the risks and its overall responsibility for all ICT-related functions. Institutions must also make an effort to learn from ICT incidents by ensuring that they have staff capable of gathering information and conducting reviews after ICT-related incidents. Those reviews should focus on matters such as the promptness of the response, the quality and speed of analysis, and the effectiveness of communication. 

A common thread running through these expectations and observations is the availability and quality of data. As the Central Bank of Ireland remarks, ''[w]ithout robust validation and reconciliation of data sets, particularly where this data drives AI system generated actions and decisions, there is a heightened risk that strategic, financial or operational actions will be based on incomplete data, potentially leading to poor outcomes for both the firm and its investors.'' In order to make informed decisions at all levels of the organisation, including at board level, and to support strategic, financial and operational actions that are increasingly reliant on AI, institutions must closely examine the quality of their data streams. 

Concentration risk 

Institutions should focus not only on the resilience of their own systems and processes, but also on resilience from a broader, systemic perspective. This point was starkly illustrated on 20 October 2026, when a major Google Cloud outage disrupted the operations of institutions across the UK's financial sector. The consequences were far-reaching: affected institutions experienced disruptions of sufficient severity to make ATM withdrawals unavailable to their customers for the whole day. Given the potentially severe consequences, institutions should incorporate into their comprehensive action plan any identified dependencies on external factors such as risks posed by third-party providers of cyber software and solutions that they depend on. The importance of mapping these dependencies and associated risks is becoming increasingly evident. In October 2026 alone, two material incidents occurred: in addition to the Google Cloud outage, the American software company Citrix was targeted and hacked through vulnerabilities in its system. This affected several critical Dutch and other European organisations and caused others to take systems offline as a preventative measure. 

Supervisory authorities from various jurisdictions have emphasized the importance of third-party dependencies in operational resilience. The German supervisor Bafin reports that in 2025, 31% of reported attacks occurred at third-party service providers. It notes that these attacks nevertheless had "a serious impact on the respective financial entity.'' Similar statements have been made by the UK Financial Conduct Authority, the US Federal Reserve, and the Dutch Central Bank and AFM. 

Specific concerns have been expressed about cloud service providers based in the US. The DNB fears that this reliance leaves institutions in a vulnerable position, as non-European service providers may act on the orders of state actors to selectively interrupt or reduce their service levels. ESMA shares this concern, warning that the financial sector risks becoming dependent on foreign regulatory regimes and susceptible to geopolitical events originating outside the EU. 

When entering into a contract with a critical or important third-party ICT service provider, DORA already requires institutions to consider concentration risk. Specifically, institutions should consider whether the contractor can easily be substituted, and whether it has entered into multiple contracts with the same or with related parties. In addition to these firm yet open-ended requirements under DORA, supervisory authorities have set out further expectations regarding how institutions can manage concentration risks. Acknowledging that reducing dependency on US providers in the short term is not feasible, the Dutch Central Bank advises institutions to develop concrete plans and scenarios for interruptions and to start using encryption to protect data. In the long term, the Dutch Central Bank would prefer institutions to reduce their dependency on US providers. 

To meet these expectations, institutions should consider the short- and longer-term consequences of incidents as well as possible reactions, including the potential impact on future cooperation with the third-party provider, any audit or access limitations that could hinder an effective investigation in the event of an incident, and the implications of potential civil liability. For financial institutions, DORA incorporates the periodical review and evaluation of third-party arrangements into their ICT third-party risk strategy. Institutions must monitor the compliance with contractual arrangements by critical or important ICT service providers, including the extent to which they realise the unrestricted rights of access. How a third party responds, cooperates and practises transparency during and after an incident should feed into the evaluation process. Additionally, the role of a third-party provider might come up during the post-incident review, which DORA requires to be carried out after a major ICT-related incident disrupts their core activities. 

As the need for a sound regulatory framework to address third-party risks becomes increasingly apparent, the Dutch Central Bank and the AFM are taking a proactive approach by advocating for additional and coordinated measures, including establishing a cross-sectoral European cloud supervisor. This would strengthen DORA's effectiveness in managing geopolitical risks and ensuring rigorous enforcement.

Unfortunately, concrete guidance on how to deal with concentration risks is scarce, and what does exist proves challenging to implement in practice. For example, where institutions are asked to migrate to European-owned software solutions. While the case for a more autonomous Europe is pressing, finding the right match among European software providers can be difficult. This gap extends to the field of AI, where European AI labs are unable to compete with their American and Chinese counterparts in terms of scale and pricing. This leaves Europe with limited capacity to build data centres and develop credible, sovereign AI solutions.

Use of AI in combatting fraud and money laundering and countering the financing of terrorism

The financial sector, particularly banking, is inherently high risk in terms of fraud, money laundering and terrorism financing. The advancement of AI affects this area in two ways. On the one hand, as the Central Bank of Ireland notes, "international fraud networks are becoming more sophisticated, taking advantage of social media platforms and increasingly using AI tools to evade bank controls and exploit customers, including those in vulnerable circumstances". Frank Elderson, Member of ECB Executive board illustrated the evolving AI threats by describing how criminals use AI-generated identities to create thousands of fake customers in order to obtain loans, causing millions of euros in losses for the bank concerned. 

On the other hand, this advancement in fraud methods warrants financial institutions to make use of AI tools to enforce policies to combat fraud, money laundering and CFT. The ECB recently ran some workshops with banks that are using AI in credit scoring and fraud detection. These banks assess they are equipped to reap the benefits of AI while prudently managing the risks, building on their existing governance frameworks and risk modelling capabilities. 

However, various challenges remain for financial institutions in the use of AI models to combat fraud and money laundering and counter the financing of terrorism. The ESRB, among others, has warned that current patching and resilience processes are built for a slower threat environment and thus are becoming insufficient. In addition, there remains a tension between privacy concerns and proportionality. As early as 2003, the Dutch Data Protection Authority found that automated, continuous and centrally pooled transaction monitoring of all customers constitutes a serious interference with the fundamental right to data protection, particularly given that only a small percentage of transactions relate to money laundering. The tension lies in the fact that any AI system depends on large-scale, pooled transaction data to work well.

Financial institutions must be able to leverage AI developments in order to respond to the use of AI tools by fraudsters. The IMF notes that when attackers operate at machine speed, defenders must do the same. In addition, financial institutions should incorporate AI cyber and ICT risks scenarios in tests and exercises, according to the FSB. Finally, financial institutions should take a proactive approach that goes beyond compliance in order to foster customer trust and confidence reiterated by the FSB, the ECB, ESAs and the Central Bank of Ireland.

Keeping up with the pace

Technological innovations and AI in particular have fundamentally changed not only how financial institutions operate, but also the environment in which they do so. The legal framework is to a large extent open-ended, and existing norms must be reinterpreted for modern applications. As legislative changes cannot keep up with the pace of technological progress, supervisory expectations expressed by national, European and global supervisors play an important role by shaping existing requirements. Such expectations underscore the need for institutions to devise understandable, explainable and manageable systems and processes that allow institutions to profit from, and grow by, using these technologies while mitigating the risks they pose. 

Meanwhile, validation, evaluation and revision of existing arrangements should occur on a constant basis, especially where third parties are involved. While supervisors are still in an exploratory, forward-looking phase when it comes to supervising these practices, institutions should be mindful that DORA already offers various possibilities to translate these expectations into hard, enforceable requirements. But more than the risk of supervisory enforcement, it is the everyday exposure to cybersecurity incidents that should keep institutions on their guard.