ECB - European Central Bank

11/16/2026 | Press release | Distributed by Public on 11/16/2026 03:11

Decoding the cyber-AI nexus

Prepared by Sándor Gardó, Benjamin Klaus, Guido Pistorius and Jonas Wendelborn

The rapid evolution of artificial intelligence (AI) is reshaping the cyber-threat landscape, with significant implications for financial stability in the euro area. This special feature explores the complex interplay between AI and cybersecurity, highlighting how advanced AI models are enabling attackers and defenders alike to enhance their capabilities. Highly capable and agentic AI systems are increasing the scale, speed and sophistication of attacks while also introducing new vulnerabilities. These include AI models themselves being exposed to novel types of attack and AI agents failing to act as their human operators intend. Critically, new AI capabilities could favour attackers, as the availability of low-cost AI agents lowers entry barriers. Meanwhile defenders remain constrained by the cost of deploying frontier models, governance challenges and slower, regulation-bound decision-making. This leaves defenders structurally disadvantaged, at least in the near term. Euro area banks, despite their high level of cybersecurity maturity, face asymmetries in access to cutting-edge AI models and rely heavily on concentrated third-party providers. This increases the risk of systemic disruption, as AI adoption transforms the channels through which cyber incidents propagate. Mitigating these risks requires banks, AI developers, cybersecurity firms and regulators to coordinate their efforts to enhance resilience, improve governance and address emerging systemic vulnerabilities.

1 Introduction

Cyber risks are of mounting concern in a rapidly evolving threat landscape. As digitalisation increasingly transforms business processes and daily life in general, there is a growing risk that failures of/attacks on underlying information and communications technology (ICT) systems will lead to disruptions. The number of publicly disclosed cyberattacks has risen considerably in recent years, including in the euro area (Chart B.1, panel a). The financial sector is a prominent - although not leading - target (Chart B.1, panel b). The threat actors behind these attacks range from individuals and organised criminals to highly sophisticated state actors. The reason for cyberattacks and their objectives also vary widely and include activist agendas, financial gain, espionage, physical damage and operational disruption. In addition, geopolitics also play an increasingly prominent role, amplifying hybrid threats and cyberattacks linked to ongoing conflicts and geopolitical tensions (Chart B.1, panel c).[1] Rapid advances in artificial intelligence (AI) are now adding to the risk landscape.

Chart B.1

Cyberattacks are becoming more frequent, increasingly targeting the euro area and critical sectors, and are often initiated from within a small set of countries

a) Number of reported global cyberattacks per annum and euro area share

b) Top target sectors of euro area cyberattacks

c) Attributed technical geographical origin of euro area cyberattacks

(2012-26; number of attacks, percentages)

(2000-26, number of attacks)

(2000-26, number of attacks)

Sources: EuRepoC and ECB calculations.
Notes: Data were retrieved on 10 November 2026. Following EuRepoC methodology, cyberattacks are included if they compromise the confidentiality, integrity or availability of information and information systems, have been publicly reported, possess a political dimension or target critical infrastructure. This deliberately excludes certain incidents, such as many financially motivated ransomware attacks against commercial entities, when they affect specific stakeholders but receive limited attention from political actors. Panel a: the European Repository of Cyber Incidents (EuRepoC) expanded its inclusion criteria as of 1 February 2023, which potentially explains the spike in attacks from this period onwards. Cyberattacks are recorded with lags. The dots show extrapolated values for the full year 2026 based on actual attacks recorded.

Cyber risks arise out of a complex interplay of factors, all of which need to be distinguished conceptually from each other. In its simplest form, cybersecurity may be defined as the preservation of confidentiality, integrity and availability (also known as the CIA triad) of information and/or information systems through the cyber medium.[2] This generally implies that information and information systems are inaccessible to unauthorised individuals (confidentiality), are accurate and complete (integrity) and are, on demand, accessible to and usable by an authorised entity (availability). Any cyber event, meaning an observable occurrence in an information system, that compromises any of these properties constitutes a cyber incident, regardless of whether it stems from malicious activity or not. By contrast, a cyberattack is a malicious attempt to exploit vulnerabilities to damage, disrupt or gain unauthorised access to assets - be they tangible or intangible - that hold value and require protection (Figure B.1). Such assets include people, finances, information, infrastructure or reputation. Cyber risk refers to a combination of the probability of cyber incidents occurring and the severity of their impact. Hence, the concept of cyber risk spans a wide range of circumstances and factors. While cybersecurity in practice requires a granular view of those factors, policy-relevant analysis can rely on a higher level of abstraction.

Figure B.1

Conceptual overview of cybersecurity, events, incidents and attacks

Source: Authors' visualisation based on FSB*.
*) "Cyber Lexicon - Updated in 2023", Financial Stability Board, 13 April 2023.

Cyber risks have the potential to pose a systemic threat to financial stability. A range of frameworks have been developed to better understand how a cyber incident could turn into a systemic event.[3] Frameworks typically combine established insights into how shocks propagate through the financial system with mechanisms specific to the circumstances of a cyber incident. Systemic events generally occur when either a single highly critical entity or a large number of non-critical entities are affected. Even if a cyber incident initially affects only a few entities, its impact can amplify and propagate through the financial system via operational, financial and confidence channels. Once the absorption capacity of the system is exhausted, the cyber incident turns into a systemic event, impairing the provision of key economic functions. In addition, substitutability, risk correlation and interconnectedness play a key role. For instance, an incident at a critical infrastructure that cannot easily be replaced, or an incident at one service that reveals vulnerabilities in another, could lead to system-wide stress. As a result, the initial cyber incident may not necessarily occur within the financial system: it could also emanate from the (digital) supply chain, critical infrastructures or the broader real economy. Rapid advances in AI are now raising concerns that such scenarios are becoming more likely.[4] Against this backdrop, this special feature explores the ways in which recent advancements in AI have fundamentally reshaped the cyber threat landscape, outlines their implications for euro area financial stability and concludes with related policy considerations.

2 The cyber-AI nexus

The cyber capabilities of AI models are advancing rapidly, led by frontier models developed in the United States. Frontier models are the most advanced, general-purpose AI systems at any given moment. While US frontier models seem to set the pace in AI capability advancement, the best Chinese models tend to catch up in just a few months, while European models progress at a comparatively more measured pace (Chart B.2, panel a). It had long been anticipated that AI models would achieve advanced cyber capabilities, but the April 2026 test results for Anthropic's Mythos model, the latest frontier model at the time, indicated that this had occurred faster than the AI and cybersecurity community had expected. Initial access restrictions were therefore imposed upon Mythos. On simpler cyber benchmarks, success rates are now high across many models, including Chinese models that matched a Mythos-level performance just four months after its release (Chart B.2, panel b). By contrast, the more challenging benchmarks remain the domain of frontier models for now. This gap is expected to narrow quickly as the next generation of frontier models further extends its capabilities.

Chart B.2

As the cyber capabilities of AI models advance rapidly, usage costs vary significantly across different models and countries

a) Frontier AI models' overall capabilities over time

b) Cyber capabilities of AI models over time

c) Capabilities and usage costs of AI models

(1 Jan. 2024-10 Nov. 2026, Epoch Capabilities Index)

(1 May 2025-23 Sep. 2026, percentages)

(x-axis: Epoch Capabilities Index, y-axis: US dollars)

Sources: Ho et al.*, Wang et al.**, Epoch.ai and Openrouter.ai.
Notes: Panel a: the Epoch Capabilities Index is a composite metric that combines scores from 50+ AI benchmarks into a single general-capability scale, using the different models' relative performance across overlapping benchmarks to weight difficulty and enable comparisons, even as individual benchmarks saturate over time (see Ho et al.*). Panel b: the success rate measures the share of real-world cybersecurity vulnerabilities that AI agents successfully exploit (for more details, see Wang et al.**). Data extracted on 11 November 2026. Panel c: the cost refers to Openrouter.ai median cost for a 10-49 turns coding-agent session over the last 30 days. Data extracted on 27 August 2026, GPT-6 Astra added separately on 21 September 2026 and Claude Opus 5.5 on 23 October 2026. Note that the charts use different samples.
*) Ho, A., Denain, J.-S., Atanasov, D., Albanie, S. and Shah, R., "A Rosetta Stone for AI Benchmarks", arXiv, 28 November 2025.
**) Wang, Z., Shi, T., He, J., Cai, M., Zhang, J. and Song, D., "CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale", arXiv, 3 June 2025.

Leading AI models are being developed predominantly in the United States and China, albeit following different approaches. US models are typically closed-weight, proprietary systems accessible only through their vendors. By contrast, Chinese models are mostly open-weight, allowing free download and deployment on private infrastructure. This offers users two key advantages. First, once downloaded, the models remain accessible without restrictions. Second, with the weights available, users can easily refine and retrain the models. While a model's capabilities are crucial for users, access costs are important too. On average, Chinese models are much cheaper than US models while the gap in capabilities is narrowing rapidly (Chart B.2, panel c). This makes them an attractive choice for high-volume workloads or for tasks that do not require frontier capabilities, prompting some US model providers to lower their prices. Given the rapid adoption of AI, its importance in global trade and financial markets, and its transformative impact on societies and economies, the race for AI-leadership carries critical geopolitical and geoeconomic dimensions. Ultimately, the dominant global supplier of AI models may not only generate significant financial benefits but could also shape the governance standards for these models, including the presence of critical safeguards.

AI's rapid adoption and expanding use cases come with inherent risks. While AI is still relatively new, it is clearly a transformative technology with great potential to have a positive impact across various domains. As societies adapt and explore its applications, however, there is a risk that the rapid deployment of increasingly powerful models will outpace the ability of many stakeholders to keep pace. As the global use - and misuse - of AI grows, the associated risks are also becoming increasingly apparent. In line with general adoption trends, the number of publicly disclosed AI incidents and hazards has grown significantly in recent years (Chart B.3, panel a). Events of this type extend beyond cyber incidents and encompass breaches like privacy violations (e.g. non-consensual deepfakes) and broader societal concerns (e.g. the development of autonomous weapon systems).

The rise of agentic AI is driving autonomous action by AI systems, reducing human oversight and at times resulting in unintended cyber incidents. AI agents are systems that leverage AI models to make decisions, take action and use tools to pursue specific goals. Agents combine the reasoning of the underlying model with writing and running software or operating other tools, often with some degree of autonomy. Agents can be configured to achieve specific goals and can collaborate to solve complex tasks. While central to AI-driven productivity gains, agentic AI also carries significant risks of unintended consequences. A notable example is the July 2026 Hugging Face incident, where an unreleased, guardrail-free AI model identified a previously unknown vulnerability in its test environment. The model then exploited this vulnerability to access the internet and hacked into the Hugging Face software platform in search of a solution to the task it had been assigned.[5] In fact, reported AI incidents are regularly tied to high levels of AI autonomy, with humans out of the loop and a notable concentration in AI security incidents, meaning those AI incidents where an AI system is involved in cybersecurity-relevant harm or risk (Chart B.3, panel b). Such incidents illustrate both the capabilities of AI models and the need for proper guardrails to ensure their safe deployment in legitimate use cases.

Chart B.3

The number of AI-related incidents is rising, with many demonstrating high levels of autonomous action

a) AI-related incidents and security incidents over time

b) Level of autonomy comparison in AI incidents

(Jan. 2020-Oct. 2026, number of incidents)

(Jan. 2020-Oct. 2026, percentage shares of incidents)

Sources: OECD* and ECB calculations.
Notes: "AI incidents" includes all OECD AI incidents and hazards. An AI incident is defined as an event in which the development, use or malfunction of an AI system results in harm to health, critical infrastructure, human/labour/intellectual property rights, property, communities or the environment. An AI hazard is defined as an event in which such harm is plausible but has not yet occurred.** "AI security incidents" covers incidents where an AI system is involved in cybersecurity-relevant harm or risk, such as unauthorised access, hacking, containment failure, misuse of autonomous AI capabilities, or other AI-enabled cyber or system-security threats. Data were retrieved on 7 November 2026. Panel b: for the sake of this analysis, 392 incidents are identified as OECD AI incident records that have similar taxonomy labels to the Hugging Face incident. The action autonomy level include (i) no-action autonomy: the system cannot act and the human uses or disregards AI system's recommendations or output at will, (ii) low-action autonomy: the system can act if the human agrees, (iii) medium-action autonomy: the system can act unless the human vetoes, and (iv) high-action autonomy: the system acts without human involvement.***
*) "AIM: AI Incidents and Hazards Monitor", OECD, 2026.
**) See "Defining AI Incidents and Related Terms", OECD Artificial Intelligence Papers, No 16, OECD, 6 May 2024.
***) See "OECD Framework for the Classification of AI Systems", OECD Digital Economy Papers, No 323, OECD, February 2022.

Improving capabilities, the race for increasingly powerful models and the rise of agentic AI are shaping cybersecurity dynamics for attackers and defenders alike. Rising cyber threats were a growing concern even before AI tools became widely accessible, and automated attacks, like self-propagating malware, are not fundamentally new. However, leveraging AI enables threat actors to significantly enhance the scale, speed and sophistication of attacks. Evidence from an AI lab and security researchers suggests that AI tools are being deployed across the entire attack chain, ranging from target selection (reconnaissance) and vulnerability identification (possibly involving social engineering) to exploiting vulnerabilities and navigating within the target system to achieve objectives.[6] Recent incidents involving financial institutions show that attackers can circumvent sophisticated security arrangements and reach systems that move or control assets by compromising trusted identities, manipulating authorised workflows or exploiting third-party dependencies without breaking cryptographic protections or exploiting underlying financial protocols. With agentic AI, such steps can be executed autonomously, with two key implications. First, barriers to entry have been reduced significantly. Attacks that once required extensive specialisation and domain-specific knowledge of target systems are now feasible for less sophisticated attackers. Second, the speed and scale of attacks are limited more by the financial resources of the attackers and less by their human expertise. At the same time, defenders also benefit. From a preventive standpoint, AI tools can help to reveal vulnerabilities, create patches and improve penetration testing. Also, AI tools can be integrated into defensive systems to ensure the activity monitoring and incident response that is increasingly needed to keep up with the speed of agentic attacks.

Chart B.4

As cybersecurity vulnerabilities are discovered at an increasing rate, both exploitation and patching are moving faster, indicating the dual-use nature of AI

a) Cybersecurity vulnerabilities discovery over time

b) Exploitation and the share of zero-day cases of cybersecurity vulnerabilities

c) Microsoft's cybersecurity vulnerability patching

(Jan. 2015-Oct. 2026, number of vulnerabilities)

(2018-26; number of exploitations, percentages)

(June 2024-Oct. 2026; number of patches, days)

Sources: CVE.org, National Institute of Standards and Technology, Cybersecurity & Infrastructure Security Agency, VulnCheck, Microsoft Security Response Center (MSRC) and ECB calculations.
Notes: Data extracted on 2 November 2026. Panel a: cybersecurity vulnerabilities are commonly referred to as common vulnerabilities and exposures (CVEs), forming part of an industry-standard dictionary and catalogue of publicly disclosed vulnerabilities. The vulnerability severity classification ranges from low, medium or high to critical. Panel b: zero-day exploitation is defined as a cybersecurity vulnerability exploitation occurring on the same day or before public CVE disclosure. The methodology is inspired by the Zero Day Clock; the discrepancy is due to differences in data coverage and timestamp availability. Panel c: "Time to patch" is defined as the time elapsed between the CVE reservation date and the corresponding patch release date for each patched CVE. The reservation date is when a unique CVE is assigned and used as a proxy for when the vulnerability is identified. The monthly median is presented as a three-month moving average.

Faster vulnerability discovery enables both faster exploitation and quicker patching, highlighting the dual-use nature of AI tools. The number of publicly disclosed vulnerabilities has surged with the widespread availability of AI models, with a notable spike following the release of highly cyber-capable models to trusted partners in 2026 (Chart B.4, panel a).[7] Close to half of these vulnerabilities are classified as critical or highly severe. The growing volume of new vulnerability disclosures means it is not feasible to conduct a comprehensive human review of all submissions. This has led to governance challenges and a higher risk of false positives in vulnerability databases.[8] While vulnerabilities are systematically documented, data on confirmed exploits remain scarce, likely due to a large grey area of undiscovered or undisclosed cyberattacks. Available data show a rising share of zero-day cases over the past few years, where vulnerabilities are exploited before or on the day of disclosure (Chart B.4, panel b). While this share has stabilised recently, it points to a narrowing time window between discovery and potential exploitation. Meanwhile, major software providers have released a record number of patches in significantly shorter time frames thanks to highly cyber-capable AI models (Chart B.4, panel c). Software firms are also exploring AI-assisted formal verification to mathematically prove specific security properties of software, an approach historically constrained by the time and expertise required.[9] Operational challenges remain, however: not only must patches be developed, they must also be deployed by end users. Once a patch is released, attackers can reverse engineer the vulnerability and can actively target systems that have yet to apply it. In addition, the rushed development of patches could result in new vulnerabilities or instabilities. This implies there is less time to deploy patches, requiring organisations to balance the urgency of patching against the potential disruption of operations during patching and the limited time available to test patches prior to their deployment.

Defenders also need to mitigate the risks associated with AI usage within their organisations, including the expanded attack surface created by such usage. AI-related cyber incidents may arise due to AI agents being insufficiently supervised, being granted excessive privileges or performing actions not intended by their human operators. This issue, commonly referred to as (mis)alignment, could lead to illegal or unethical behaviour. AI usage also significantly increases the attack surface defenders need to address, as AI is vulnerable to new types of attacks (Chart B.5, panel a). For example, AI models can be instructed to perform malicious actions through prompt injections, which are difficult to intercept. In addition, concerns have recently emerged about potential backdoors in open-weight models, such as unprompted uploads of information on remote servers.[10] This could pose security risks even when such models are deployed on private infrastructure, particularly if granted access to tools and internal systems. At the same time, the supply chain supporting AI tools may expose organisations to further cybersecurity risks.

Improving guardrails for AI models is crucial but will only partly mitigate risks, as threat actors constantly seek ways to circumvent them. As organisations gain more experience of AI-related cyber incidents, both built-in guardrails and controls around AI agents will improve, enabling safer deployment in corporate environments.[11] Given the business-like ecosystem of cybercrime, however, specialised services are emerging that offer so-called "jailbreaks" that circumvent the guardrails of legitimate AI models. While the guardrails of closed models may not always work as intended, open-weight models can be stripped of guardrails entirely, which is done mostly through a process called "abliteration" (Chart B.5, panel b).[12] This can help defenders use AI models to find vulnerabilities but, due to the dual-use nature of the technology, it also enables offenders to exploit these vulnerabilities. Abliterated models are, to a large extent, built on Chinese models. Most of them are small enough to run on (high-end) consumer devices and thus can be run at low cost on attackers' hardware, beyond the oversight of any provider (Chart B.5, panel c). In addition, commercial platforms are emerging that also offer access. As such, even if improved guardrails make jailbreaking more challenging, abliterated open-weight models will continue to fill the gap left by restricted systems.

Chart B.5

AI models are vulnerable to specific attack types and could be modified to bypass built-in guardrails designed to limit their use in offensive cyber operations

a) Concentric threat flow for the top ten LLM vulnerabilities

b) Number of guardrail-free models

c) Characteristics of guardrail-free models

(Jan. 2024-Oct. 2026; number, percentages)

(Jan.-Oct. 2026, percentages)

Sources: OWASP, Hugging Face and ECB calculations.
Notes: Panel a: adapted and simplified from OWASP*. Attack flows inwards, defence pushes outwards. Some vulnerabilities can sit in more than one layer. Panels b and c: analysis is based on models available on the Hugging Face software platform. LLM stands for large language model. Panel c: parameter size refers to the total number of trainable internal variables (weights and biases) inside a neural network, which serves as a metric for the model's overall capacity, complexity and computational scale.
*) "OWASP GenAI LLM Top 10 2026", OWASP GenAI Security Project, 3 August 2026.

Using AI tools for defensive purposes comes with a number of challenges and trade-offs which could further exacerbate the imbalance between attackers and defenders. First, the surge of new discoveries is forcing defenders to triage which vulnerabilities their patching efforts should focus on. This is reinforcing the dilemma faced by defenders: while attackers only need to exploit a single vulnerability, defenders must strive to fix all vulnerabilities. Second, using frontier models for system defence might be a problem. Such models could be too costly, both financially and in terms of latency.[13] While their advanced reasoning capabilities could help to reveal attempted intrusion or other unusual behaviour, they would likely introduce system slowdowns that would impede business operations. Organisations might also be reluctant to share security-relevant information with models hosted on external servers. Standard models pose the risk that guardrails designed to prevent them being used for cyberattacks could hinder legitimate defensive action, potentially leaving defenders vulnerable at critical moments.[14] Given all this, open-weight AI models that defenders can run and control on their own infrastructure are likely to be key to enhancing defensive capabilities. Third, the threat landscape requires defenders to use AI tools, yet these tools could be vulnerable to attack and require robust governance, which remains underdeveloped as AI adoption is still at a relatively early stage. These challenges are compounded by the fact that cyber defence is often bound by lengthy decision-making processes, guidelines or regulations, putting defenders at a structural disadvantage.

Even if defenders succeed in making systems significantly more secure, cyber risks will still persist. Essentially, there are two possible scenarios for how the AI-powered race between attackers and defenders might unfold. In one scenario, the current situation persists, with new AI models continually uncovering new vulnerabilities and exploits. In the other scenario, there is a convergence towards significantly safer systems as legacy vulnerabilities are progressively patched and adopting "security by design" principles during development leads to software becoming more and more impenetrable. Many AI labs are working to enable the latter scenario by granting trusted cybersecurity partners advanced access to their latest models, with cyber guardrails disabled. Even if these efforts succeed, however, cyber risks will persist as threat actors adapt. For example, attackers could use AI tools to amplify the scale, speed and effectiveness of social engineering attacks to gain system access via legitimate users. In addition, the geographical distribution of vulnerabilities could be impacted by geopolitical fragmentation and resulting access restrictions to AI models.

While cyber risks are not fundamentally new, advancements in AI come with two key risks. First, there is a tangible risk that cybersecurity teams and systems may become overwhelmed by the sheer volume of issues they have to address, leaving gaps that attackers could exploit. This risk is particularly acute during the transitional period, as AI-related risks continue to emerge and corresponding solutions and governance frameworks are still being developed. The challenge is compounded by the fact that defenders are not a homogeneous group but comprise security teams, software developers, open-source maintainers, researchers, consultants or IT staff within individual organisations. Second, AI has significantly altered the distribution of capabilities across threat actors. In the past, the most advanced and resourceful attackers were typically state or state-affiliated actors capable of executing high-impact attacks. However, these actors operate within the broader balance of power between nations, where national interests, geopolitical strategies and credible deterrence shape actions. Agentic AI now allows a wider range of threat actors to carry out high-impact attacks on a wider range of targets, such as critical infrastructure. This elevates the risk of systemic disruption caused by "rogue" actors.

3 Implications for the euro area banking sector

Overall, banks are well aware of cyber risks and are increasingly acknowledging the additional challenges posed by AI. Compared with other industries, the financial sector demonstrates a relatively high level of maturity when it comes to cybersecurity.[15] In fact, survey results show that cyber-related operational risks have been a top priority for risk officers in the financial sector for many years now (Chart B.6, panel a). Although banks do not typically communicate extensively on cyber risks, more and more of them are addressing the topic in their earnings calls (Chart B.6, panel b). Some banks have explicitly highlighted the emerging risks associated with frontier AI models and emphasise that cybersecurity is a high-priority investment (Chart B.6, panel c). In addition to potential AI-enabled cyberattacks, banks face risks from their internal use of AI, including incidents arising from weak governance or issues with model alignment.

Chart B.6

Global banks are alert to cyber risks and are communicating more on the topic

a) Rank of cyber-related risks in top ten operational risks of global financial firms

b) Share of global banks mentioning cyber risks in earnings calls

c) Types of cyber threat discussed in global banks' earnings calls

(2018-26, survey rank)

(2022-26, percentages)

(2022-26, total)

Sources: Risk.net, earnings call transcripts and ECB calculations.
Notes: Panel a: data are based on a survey of financial firms identifying their ten most significant operational risk concerns for the year ahead. Panels b and c: analysis is based on the earnings call transcripts for 82 banks in the Stoxx Global 1800 Banks Index.

Cyber incidents at euro area banks have not been systemic to date. There are only a few real-life examples of major cyber crises (such as the 2017 NotPetya attacks), and the euro area banking system has not experienced a systemic cyber event to date.[16] Data collected under the EU's Digital Operational Resilience Act (DORA) show that malicious actions account for a relatively small share of major cyber incidents, with most incidents having limited impact (Chart B.7).[17]

Chart B.7

Major cyber incidents at euro area banks are rarely linked to malicious actions, while failures at third-party providers account for a significant share

Breakdown of high-level root causes triggering a major incident

(Q1 2025-Q3 2026, percentages)

Sources: ECB (DORA) and ECB calculations.

There are three key asymmetries that may make the euro area banking system vulnerable to the AI-enabled increase in cyber risks.[18] The first concerns countries with and countries without access to cutting-edge frontier models. This affects the euro area as a whole, as frontier AI models are predominantly developed outside the EU and EU entities have often faced restrictions in accessing the most capable models. This could hinder their ability to identify vulnerabilities within their systems and to respond effectively to cyberattacks. The second asymmetry is between attackers and defenders, with defenders currently at a structural disadvantage, as discussed above. This is a concern for organisations worldwide. The third asymmetry regards banks equipped with a sophisticated cybersecurity apparatus and those with more limited resources. This may be down to fixed cost and resource constraints and could pose a risk, as the weakest link in a network may affect the stability of the entire system. A survey conducted by the European Banking Authority showed that 56% of banks did not face a cyberattack in the first half of 2026 that either resulted or could have resulted in a major ICT incident, up from 49% in the previous half.[19] At the same time, 11% of banks saw 11 or more attacks, up from 7% previously. These results may be early indications of a bifurcation dynamic starting to play out between different sets of banks. All these asymmetries create structural cybersecurity disadvantages for euro area banks whose relative financial strength, at the same time, may make them attractive targets for cyber criminals.

Heavy reliance on third-party service providers could serve as a mitigating factor but it also poses additional risks with possible systemic implications. Euro area banks rely heavily on third-party service providers, particularly for their ICT systems. These dependencies are highly concentrated among a small number of large providers from outside the EU. On the one hand, outsourcing to such providers can enhance resilience, as they often have privileged access to highly cyber-capable frontier models and possess the expertise to make them work. These large providers are therefore well positioned to safeguard both their own systems and those of their clients. On the other hand, they face the same cybersecurity challenges posed by frontier models as everybody else, thereby increasing the attack surface for banks that rely on them. Strong concentration among third-party providers could create single points of failure and reduce substitutability as a mitigating measure. Thus, a cyber incident at a third-party provider could simultaneously affect large parts of the euro area banking system. Currently, around one-fifth of cyber incidents affecting euro area banks originate from third-party providers (Chart B.7). Also, outsourcing to non-EU providers creates critical dependencies that could expose euro area banks to risks emerging from geoeconomic fragmentation. Just as certain frontier AI models have been placed under export controls, in an extreme scenario similar restrictions could be applied to other software or services, potentially causing significant disruptions if they form a critical part of ICT systems and cannot be replaced quickly. This risk could be reinforced by lock-in effects if these providers are offering the most cyber-secure solutions due to their privileged access to frontier models. However, given the high costs such restrictions would inflict on affected software and service providers, this scenario is not especially likely in current circumstances.

AI not only increases cyber risk directly but may also amplify the mechanisms that could escalate a cyber incident into a systemic event. As mentioned above, agentic AI has the potential to significantly increase the sophistication, scale and speed of cyberattacks. This capability makes it possible for attackers to target a larger number of financial institutions simultaneously with fewer resources, thereby increasing the likelihood of systemic disruptions. Also, AI could enhance or specifically exploit amplification channels. A cyberattack might, for example, be paired with a targeted misinformation campaign on social media (automated and scaled using AI) to amplify the attack's impact by undermining customer or counterparty confidence. Moreover, growing AI-automation within financial institutions could further result in operational vulnerabilities, as certain processes could become harder to perform manually should an incident affect AI's ability to execute. Similarly, AI adoption could alter the network of financial interlinkages or the speed of market interactions, potentially increasing the speed or scope of contagion via direct exposures or financial interlinkages. Whether AI enhances amplification or, instead, increases reaction agility will depend on the nature of a cyber incident.

4 Conclusions

Rapid advances in the cyber capabilities of AI models are contributing to heightened cyber risks. These risks stem from the deployment of agentic attacks, an expanded attack surface and the potential for AI agents to act in ways not intended by their operators. While defenders also benefit from AI tools, they may face a structural disadvantage, at least in the short term. Looking ahead, two key risks stand out. First, in the near to medium term, the pace of AI-powered advancements in cyber capabilities could outstrip the ability of organisations to adapt, with the risk of defences becoming overwhelmed. Second, as AI models continue to evolve, the capabilities of different threat actors converge towards higher levels of sophistication, increasing the probability of more severe impacts. These risks are further exacerbated by geopolitical imbalances and competition over who should develop capable AI models, who should gain access to them and who should set the governance standards for their use. In particular, geoeconomically motivated restrictions to or the withdrawal of access to the most capable AI models or related critical ICT services could result in significant operational challenges.

Effective cyber risk mitigation requires coordinated efforts and decisive action. As AI proliferates and the risk of agentic cyberattacks increases, swift action is essential to prevent (or at least mitigate) the emergence of systemic cyber risks to euro area financial stability. Where not done already, adapting cyber defences at the level of individual firms is clearly crucial. This could include adopting principles of defence in depths and assumed compromise, heightened monitoring of any traffic with AI models and promoting awareness and a speak-up culture to build resilience to social engineering. However, addressing these challenges only at the level of individual institutions would fall short of achieving this objective. Instead, a coordinated effort is needed between banks, AI developers, software and ICT providers, cybersecurity firms, open-source maintainers, and national and international authorities. This entails enhancing operational and cyber resilience at the firm level, as well as collective contingency planning and corresponding frequent testing of established contingency mechanisms. Cyber mapping could help to identify common exposures and single points of failure in ICT supply chains. Raising societal awareness and preparedness (e.g. holding sufficient cash) is equally important to strengthen resilience more broadly. In this regard, EU policy initiatives like the AI Act and DORA already play an important role in addressing AI-related risks as well as third-party dependencies and concentration risks. Besides ensuring the effective implementation of existing frameworks, authorities need to continue their efforts on strengthening coordination between stakeholders and further developing European crisis management structures. The continuous evaluation of emerging gaps in both crisis prevention and management frameworks will be the key to retaining agility and keeping pace with technological developments.

ECB - European Central Bank published this content on November 16, 2026, and is solely responsible for the information contained herein. Distributed via Public Technologies (PUBT), unedited and unaltered, on November 16, 2026 at 09:11 UTC. If you believe the information included in the content is inaccurate or outdated and requires editing or removal, please contact us at [email protected]