(1)

Leading artificial intelligence (AI) providers have developed Frontier AI Models (FAIMs), which are highly capable in the cybersecurity domain and able to carry out fully automated cyber-attacks on complex systems, including through vulnerability discovery and the development and weaponisation of exploits, thereby fundamentally changing the cyber-threat landscape.

(2)

FAIMs are significantly better than earlier AI models in finding ways to carry out cyberattacks, outperforming earlier models in cost, speed and accuracy, now rivalling leading human experts.

(3)

FAIMs have the capability to threaten the systems underpinning the information and communications technology (ICT) environments on which financial infrastructure relies, increasing the probability and severity of cyber incidents with systemic impact.

(4)

FAIMs have demonstrated the capability to discover vulnerabilities and craft novel exploits, thereby exposing major operating systems and software used across almost all organisations’ ICT environments to the risk of cyber-attacks.

(5)

Historically, the discovery of such vulnerabilities has been rare and often made by security researchers, allowing for targeted remediation. In those cases, software providers are often given a standardised timeframe – usually 90 days – to patch the vulnerability before its public disclosure. However, FAIMs are likely to increase the volume of vulnerabilities found.

(6)

Current patching practices in the financial system are predominantly reactive, relying on periodic updates and ad hoc responses to disclosed vulnerabilities. This approach works in a threat environment where time allows for vulnerability discovery to remain manageable. However, with an increase in volume of vulnerabilities, current practices may become insufficient to maintain operational resilience.

(7)

This increase in volume also exposes current patching processes to overloading, as they prioritise remediation measures based on risk and require testing of software patches before implementation to ensure operational stability. If too many vulnerabilities with critical impact are discovered within a short timeframe so that the number of critical software patches required is substantially increased, financial institutions may have to decide between leaving themselves exposed to significant cyber risks, or reducing the patch-testing requirements and thereby risking operational incidents and outages.

(8)

The crafting of weaponised exploits was previously largely done manually and took human experts days or weeks, whereas it can now be done by FAIMs in a matter of minutes or hours. This constitutes a collapse of defensive time buffers, which are needed to maintain the continuity of critical and important functions during remediation.

(9)

With the proliferation of weaponised exploits corresponding to critical vulnerabilities powered by AI and the corresponding reduction of defensive time buffers, the risk to individual critical or important functions and institutions increases sharply. If such risks materialise simultaneously across those functions and institutions, this could result in a permanent increase in systemic cyber risk for which, at present, there is no fully effective mitigation framework available. Due to the strong interconnectedness within the Union financial sector, as well as between that sector and other sectors and jurisdictions, this situation poses a systematic risk and has the potential to create systemic fragilities.

(10)

Furthermore, FAIMs are currently being developed by only a small number of AI providers, several of whom have expressed concerns that forthcoming FAIMs may be too powerful for unrestricted public access.

(11)

AI is already being used by malicious actors to enhance cyber-attacks. It is likely that such actors will eventually obtain, whether through leaks, theft, independent development, or open access, AI models with capabilities comparable to those currently deployed in controlled defensive settings.

(12)

While FAIMs may significantly increase offensive capabilities and the likelihood of sophisticated cyberattacks, they will also strengthen defensive capabilities, specifically regarding vulnerability discovery, data correlation and remedial action by individual institutions. However, in the short-to-medium term, the increase in speed, scale and accuracy of offensive capabilities are likely to outweigh the benefits.

(13)

The rapid development of FAIMs makes it necessary to address the identified risks to financial stability without delay so as to prevent their materialisation or at least to mitigate their potential impact. To ensure financial stability in the Union financial markets, all European Systemic Risk Board (ESRB) members should consider the implications of FAIMs on operational capability and resilience of financial institutions in the context of their macroprudential and microprudential policy actions. This is without prejudice to the monetary policy mandates of the central banks in the Union. As an example, and among other authorities’ current initiatives, the ECB, in its role of banking supervisor, has requested the significant institutions to assess the impact of the evolving threat landscape without delay, and to develop by 31 October 2026 a comprehensive action plan outlining concrete measures aimed at addressing such risks (4).

(14)

As AI technology proliferates, it is important to ensure that all critical infrastructure and especially financial institutions are prepared to face cyber-attacks powered by FAIMs, and make timely and appropriate preparations. However, such defensive efforts by single entities would be insufficient. An effective response and risk mitigation requires a coordinated answer involving all affected parties including AI providers, software providers, security firms, open-source maintainers, financial institutions, and authorities at both national and Union level.

(15)

When addressing the issues above, it is important to recall relevant Union legal and policy frameworks, in particular Regulation (EU) 2022/2554 of the European Parliament and of the Council (5), Regulation (EU) 2024/1689 of the European Parliament and of the Council (6), Regulation (EU) 2024/2847 of the European Parliament and of the Council (7) and Recommendation of the European Systemic Risk Board (ESRB/2021/17) (8).

(16)

Regulation (EU) 2022/2554 provides a comprehensive framework for financial institutions to identify, manage, monitor and mitigate ICT risks, including risks stemming from cyber-attacks. Given the interconnectedness of the financial sector and its reliance on ICT systems and third-party service providers, a successful cyber-attack affecting one or more financial institutions or to their critical service providers could have significant systemic implications. Financial supervisory authorities should give appropriate priority to supervisory activities in this area, having regard to the potential impact of such cyber-attacks on the stability, integrity and trustworthiness of the financial system, as well as to the operational, financial and consumer harm that may result.

(17)

Regulation (EU) 2024/1689 establishes harmonised rules for the placing on the market, putting into service and use of AI systems and general-purpose AI models in the Union, including, in certain circumstances, systems and models provided by AI providers established in third countries where the output produced by the AI system is used in the Union, or an AI model is integrated into a system placed on the Union market. This Regulation also provides specific rules, including a stricter regulatory regime, for models classified as general-purpose AI models with systemic risk.

(18)

Regulation (EU) 2024/2847 introduces mandatory cybersecurity requirements for manufacturers which cover the planning, design, development and maintenance of hardware and software products with digital elements placed on the Union market. It also requires manufacturers to handle vulnerabilities during the lifecycle of their products. Once fully applicable in December 2027, financial entities placing such products on the Union market will also need to ensure those products comply with these requirements.

(19)

Recommendation ESRB/2021/17 recommended the development, by the European Supervisory Authorities (ESAs) jointly through the Joint Committee and together with the European Central Bank (ECB), the ESRB and the relevant national authorities, of an effective Union-level coordinated response in the event of a major cross-border cyber incident or related threat that could have a systemic impact on the Union’s financial sector. This resulted in the establishment of the pan-European Systemic Cyber Incident Coordination Framework (EU-SCICF), which provides a valuable and operational platform for information exchange and coordination among financial authorities.

(20)

Against this background, the ESRB hereby adopts a warning of a general nature addressing the significant financial stability risks amplified by the development of FAIMs. The warning takes account of the analytical work in the ESRB note Addressing Frontier AI Models with cyber capabilities from a financial stability perspective published in June 2026 (9),