Summary

• European banking supervisors are pressing institutions to demonstrate that essential services can be recovered after more sophisticated cyber incidents.
• Independent recovery environments, shared technology dependencies and faster vulnerability remediation are receiving greater scrutiny.
• Existing operational resilience requirements remain in force; the latest supervisory engagement does not introduce a new AI-specific regulation.

European banking supervisors are intensifying scrutiny of financial institutions’ ability to recover from serious cyber incidents, amid concern that increasingly capable artificial intelligence could accelerate attacks and compromise the technology required to restore essential services.

The Bank of England and European Central Bank are examining banks’ recovery arrangements, including their capacity to rebuild operations using infrastructure independent of compromised systems, according to Reuters reporting published on 13 October. The engagement reflects concerns about a changing threat environment rather than the introduction of new binding rules.

Supervisors have long required banks to maintain business continuity and disaster recovery arrangements, but recent developments in AI are changing assumptions about how quickly vulnerabilities might be exploited. Systems capable of carrying out extended technical tasks with limited human intervention could allow attackers to identify weaknesses and attempt exploitation faster than existing security processes can respond.

Although similar technology can help defenders discover and remediate vulnerabilities, the growing ability to automate parts of an attack places additional demands on institutions whose operations depend on extensive software estates and external providers. Recovery planning must also account for the possibility that an attacker has compromised administrative tools, credentials or backup systems before the intrusion is detected.

Rebuilding systems that can no longer be trusted

Conventional disaster recovery can involve restoring data and restarting services from backups or an alternative computing environment. That approach becomes more difficult when the organisation cannot establish whether the infrastructure used for recovery is itself trustworthy.

In a serious compromise, restoring an apparently intact application may reintroduce malicious software or allow an attacker to regain access through credentials that remain valid. Banks therefore need procedures for establishing a trusted environment, verifying essential data and restoring services in an order that reflects their operational dependencies.

The supervisory discussions reported this week include the concept of bare-metal recovery, in which essential systems can be rebuilt on clean infrastructure rather than relying exclusively on an existing environment that may have been affected. Such approaches are technically demanding because they require dependable software images, configuration records, access controls and backups that remain usable independently of compromised systems.

The reported scrutiny does not establish that every bank has been ordered to adopt a particular bare-metal architecture. Institutions operate different systems, and the appropriate recovery design depends on their critical services, existing infrastructure and the nature of the disruption against which they must prepare.

Recovering a banking service can also be more difficult than restoring its individual applications. Payment processing, for example, depends on coordinated systems for recording transactions, establishing customer balances and exchanging information with external networks. Restoring one component before the others are ready may leave the service unavailable or create inconsistencies that must be resolved before operations resume.

These dependencies extend beyond infrastructure owned directly by the bank. Identity services, cloud platforms and specialist financial technology providers may support several critical applications, while outsourced software can introduce additional recovery requirements and contractual dependencies.

Where multiple institutions rely on the same supplier, an incident affecting that supplier could disrupt several firms simultaneously. A bank’s own recovery arrangements may be technically sound but insufficient if it cannot operate a critical service without access to an unavailable external platform.

Existing frameworks face faster threats

Within the European Union, the Digital Operational Resilience Act has applied since January 2025, establishing requirements for managing information and communication technology risk, reporting relevant incidents and overseeing technology providers. Banks must also address resilience testing and maintain appropriate arrangements for responding to operational disruption.

The ECB’s supervisory priorities for 2026–28 place operational resilience among the authority’s principal areas of attention. Its programme includes examinations of cybersecurity and third-party risk management, alongside assessments of how selected banks would cope with disruption affecting a major cloud service provider.

ECB supervisors also intend to examine institutions’ governance of AI technologies as adoption increases. That work concerns how banks manage the risks associated with their own use of AI, which is related to but distinct from the threat of attackers using more capable models against banking infrastructure.

In Britain, the Bank of England, Financial Conduct Authority and HM Treasury issued a joint statement in May urging regulated institutions to prepare for the risks created by frontier AI. The authorities expressly said that the statement reinforced existing expectations rather than introducing additional requirements.

The Bank of England’s July Financial Stability Report subsequently examined how advances in AI could increase the volume of software vulnerabilities requiring attention and shorten the time available to address them. The report also identified operational risks arising from the defensive response, particularly where organisations must deploy patches and system changes more frequently.

Rapid remediation can reduce the opportunity for exploitation, but changes implemented without sufficient testing may themselves interrupt banking services. Financial institutions consequently face the combined task of accelerating vulnerability management while retaining enough control over production systems to avoid preventable outages.

The Bank’s analysis also distinguished simulated AI cyber capabilities from demonstrated attacks against well-defended real-world institutions. Results from controlled security evaluations indicate that frontier models have become more capable, although they do not establish that autonomous attacks can reliably compromise every type of operational environment.

AI introduces additional operational dependencies

Banks adopting AI internally must also manage systems that can act across multiple applications. Autonomous agents may be given access to monitoring tools, databases or administrative functions, potentially allowing them to perform useful work without waiting for an employee to approve every individual action.

Those capabilities require controls over authority and access. An agent configured to investigate a security incident, for example, might need permission to query logs or examine system configurations, while changes to accounts, network settings or production workloads could require separate authorisation.

Keeping records of an agent’s actions allows investigators and operational teams to understand what occurred if a task produces an unexpected result. Banks also need ways to restrict automated activity when the system behaves incorrectly or encounters instructions and information it should not trust.

Techopia previously examined the gap between AI experimentation and operational deployment in banking regulatory reporting, where governance requirements become more demanding once software begins carrying out work within established business processes. Similar questions arise when agents are introduced into security operations, although the immediate risks and controls differ.

Greater reliance on AI systems can also introduce concentration risks where banks and their suppliers use the same underlying models or service providers. An interruption affecting a shared AI platform may have consequences for several institutions, particularly if that platform has become embedded within activities that previously depended on internal software.

Such developments complicate recovery planning because banks must consider both AI-enabled attacks from outside their organisations and failures involving AI systems they have chosen to deploy. The same technologies may strengthen defensive capability while creating new dependencies that institutions need to understand.

The latest supervisory engagement reflects a broader shift towards demonstrating that banks can restore critical operations under increasingly severe conditions. The underlying obligations are established, but the assumptions used to test them are changing as AI increases the pace of vulnerability discovery and places greater pressure on the systems and personnel responsible for recovery.