Short answer. The ECB set a hard date: every significant bank must submit an action plan against AI-enabled cyberattacks to its Joint Supervisory Team by 31 October 2026 (letter SSM-2026-0301, 7 July 2026). A credible plan names concrete measures, allocates resources, assigns clear roles, and sets timelines, and it builds on the bank’s existing cyber-risk strategy rather than sitting beside it. The work is an operating-model problem, not a tool-buying exercise: decide who owns each control, how it is monitored, and who signs the audit trail.
The ECB’s July 2026 letter, addressed to the CEOs of the significant institutions it supervises directly, treats AI-enabled attacks as a shift in the threat model, not a new tool category. Frontier AI models compress the time between a vulnerability being found and being exploited, so a plan written for last year’s patch cycle is already behind. Banks are expected to assess how that change affects ICT resilience and to strengthen their defenses on a fixed timeline.
In practice, the strongest action plans converge on a short list of measurable commitments rather than a strategy narrative. The recurring priorities across supervised banks are concrete:
The pattern is clear: every line needs an owner, a number, and a date. A plan that lists intentions without those three fails the same test a stalled AI pilot fails. It never reaches production.
More than 85% of banks under European banking supervision already use AI, so the ECB is regulating a live operational reality, not a forecast. The action plan does not arrive in a vacuum. It sits on top of DORA, in force since 17 January 2025, which already requires financial entities to run an internal governance and control framework for ICT risk and to report a significant cyber incident within two hours of classifying it as significant. The ECB has extended its annual IT Risk Questionnaire deadline from September 2026 to February 2027 to free capacity for exactly this work, which signals how seriously the supervisor treats the October date.
The EU AI Act adds a second clock. Under Regulation (EU) 2024/1689, AI systems used for credit scoring and insurance risk assessment are high-risk (Annex III) and carry obligations for risk management, logging, and human oversight that reach production in the coming supervisory cycles. A bank that treats the ECB cyber plan, DORA, and the AI Act as three separate binders duplicates work and creates gaps between them. Treated as one operating model, the same controls, owners, and audit trail answer all three.
Significant banks must submit an AI cyber action plan to their Joint Supervisory Team by 31 October 2026 (ECB letter SSM-2026-0301).
The letter went to the CEO, and that is deliberate. Accountability sits at the top, but the plan is authored across security, IT, third-party risk, operational resilience, risk management, and the executive board. The failure mode is a plan owned by “the CISO’s team” alone, because AI-enabled threats cross the line between security tooling and the AI systems the bank runs in production. The workable structure is a single accountable owner at board level, a named lead for each measure, and a documented sign-off chain so an auditor can see who approved what. That is the same question a DORA audit asks about any AI system in operation: who signs, who owns the exception, and who reads the audit trail.
The ECB asks for four elements: concrete measures, allocated resources, assigned roles and responsibilities, and implementation timelines. “Concrete” is the load-bearing word. A measure such as “improve monitoring” does not qualify; “deploy behavioral monitoring on all internet-facing services by Q1 2027, owned by the SOC lead, funded from the resilience budget” does. Each measure should trace back to a specific AI-enabled risk the bank has identified, so the plan reads as a response to an assessment rather than a generic checklist. Because the ECB requires the plan to build on the existing cyber-risk strategy, reuse the control library the bank already maintains under DORA and extend it, rather than starting a parallel document.
DORA governs how a bank withstands and recovers from ICT disruption; the ECB letter sharpens that toward AI-specific attack speed. The AI Act governs the AI systems the bank builds and deploys, including the defensive AI tooling named in the plan. The connective tissue is the register of third-party providers: DORA already requires it, the ECB wants AI service providers reassessed inside it, and the AI Act assigns provider and deployer duties across the same vendors. One maintained register, mapped to owners and contracts, satisfies the overlap. Maintaining three disconnected lists guarantees a gap an auditor will find.
With roughly a month before the deadline, sequence beats ambition. First, run a focused AI-threat assessment against the current control set and record the gaps; the plan is only credible if it answers a documented assessment. Second, assign a single accountable owner and a named lead per gap, because unassigned measures are the most common reason a plan reads as intent rather than commitment. Third, convert each gap into a measure with a resource line and a date, then take the whole plan through board sign-off so the submission carries the authority the ECB addressed to the CEO. A tight plan of ten owned, funded, dated measures is stronger than forty aspirations.
Submission is the start, not the finish. The ECB will read the plan as a commitment it can supervise against, so the controls need live monitoring, a refresh cadence tied to the changing threat, and an audit trail that shows measures moving from planned to in production. This is where a bank’s AI-in-production discipline pays off: the same observability, logging, and rollback practices that keep a credit-scoring model accountable also prove to a supervisor that a cyber measure is real. A plan that cannot show progress by the next supervisory touchpoint is a plan that was written for the deadline, not for the risk.
Ableneo works where this problem actually lives: getting AI systems into production and keeping them accountable inside regulated banks and insurers in Central Europe. Across its 2025 portfolio, roughly four of five AI projects reached production, and its work with financial-services clients means the governance, monitoring, and audit-trail practices the ECB now expects are the same ones Ableneo already applies to AI in operation. An action plan is an operating-model deliverable, and that is the discipline Ableneo brings: clear owners, measurable controls, and a documented chain from decision to production. You can see more of that approach at ableneo.com.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.