Frontier AI: what the ESAs expect from banks under DORA

Update – 8 November 2026. In parallel to the ESAs’ statement, the ECB wrote to the CEOs of significant institutions on 7 July 2026 on AI-enabled cybersecurity threats, asking each of them to submit an action plan to its Joint Supervisory Team by 31 October 2026. That deadline has now passed: the ECB will run a horizontal analysis of the plans received and share its conclusions, and it has moved its annual IT Risk Questionnaire to February 2027. Less significant institutions were not in scope, but should expect national supervisors to ask the same questions.

On 31 July 2026, the three European Supervisory Authorities (EBA, ESMA, EIOPA) published a joint statement on ICT risks from frontier AI models. It creates no new obligation. It does show how supervisors will read the DORA requirements already in force, which is enough to change your resilience priorities.

A convergence statement, not a new rulebook

The document (JC 2026 25) builds on DORA and the AI Act. It notes that highly capable models can speed up vulnerability discovery, target shared infrastructure and exploit single points of failure. It cites the ESRB’s warning that misuse of such models could substantially affect the financial system, and the Commission’s July 2026 Action Plan on Cybersecurity and AI.

The ECB is referenced on one specific point: significant institutions should address open supervisory findings on priority ICT areas without delay. The statement sets no deadline of its own. It does indicate that AI-related risks will be embedded in examination methods in the 2027 oversight plan.

Fourteen actions across three themes

The annex sets out fourteen risk-mitigation actions, each with examples from basic to advanced maturity.

What it means for governance

Authorities are asked to check that management bodies are genuinely engaged: clear accountability, timely response plans, sufficient investment in cyber resilience, and a risk appetite framework updated with AI-specific metrics. Oversight by the management body shifts from a periodic rhythm to more continuous decision-making, supported by automated monitoring and escalation rules.

Proportionality still applies: measures are calibrated to size, risk profile and complexity of activities.

A practical reading for a mid-sized bank: you are not expected to run the same programme as a global systemic institution, but you are expected to show that you assessed the exposure, chose measures that match it, and can explain why. Documentation of that reasoning, dated and approved, is what a supervisor will ask for first. The same logic applies to smaller entities and to groups, where the parent’s inventory and testing programme must actually reach every subsidiary and its own vendors.

Third-party risk in the foreground

The ESAs stress the supply chain: assess and enforce cybersecurity standards across service providers, hardware and software suppliers, and open-source communities. For a bank, this ties the topic directly to the DORA register of information and to concentration analysis.

Fitting it into DORA and the AI Act

The ESAs anchor the topic in four DORA pillars already in force: the ICT risk management framework, testing, incident and recovery management, and ICT third-party risk. The AI Act completes the picture: it places transparency and cybersecurity obligations on providers of general-purpose models with systemic risk. In practice, your DORA compliance review is the natural entry point. Rather than opening a parallel “AI” workstream, fold these fourteen actions into your existing testing, continuity and vendor-monitoring plans. Also check that patching and privileged-access controls, often already covered, hold up against faster exploitation.

What this changes for you

1. Extend your asset inventory to AI components. Models, APIs, datasets and AI tools supplied by vendors belong in the mapping that feeds your DORA register.
2. Add AI-assisted scenarios to your resilience tests. Exercise simultaneous failures across several systems and faster vulnerability discovery.
3. Have the management body approve AI-related risk appetite metrics. Set documented tolerance thresholds and an escalation path, and close your open supervisory findings.

Helios Advisory designs training pathways for banking and insurance professionals: see our training offer to prepare your teams for these expectations.

Sources

• EBA / ESAs, Operational resilience page and statement JC 2026 25 on frontier AI models, 31 July 2026.
• EBA / ESAs, Statement JC 2026 25 (PDF), 31 July 2026.
• European Union, Regulation (EU) 2022/2554 (DORA), EUR-Lex.