On July 31, the three European Supervisory Authorities, EBA, EIOPA and ESMA, published a joint statement on the ICT and cybersecurity risks that frontier AI models pose to the EU financial sector. Read the diagnosis first, because it is genuinely good. Frontier models could identify and exploit vulnerabilities at speed, hit weaknesses in shared infrastructure, and turn single points of failure into systemic events. AI-assisted attacks could run continuously, adapt in real time, strike multiple systems at once, and amplify supply chain compromises through automation, with cascading failures across interconnected entities. This is not scaremongering. It is a credible picture of where offensive automation is heading, and the authorities treat it as a near-term scenario, not science fiction.
Now read the therapy. Prevention: complete IT asset inventories, security by design, proactive and automated patching, least privilege, security standards for vendors. Detection: continuous vulnerability scanning, behavioural monitoring, a shift from periodic testing to near real-time verification. Management: incident response plans updated for multiple-failure scenarios, AI-enhanced scenarios in operational resilience testing. Boards must demonstrate commitment, clear accountability and adequate resources.
Every single item on that list is correct. Every single item has also been on the compliance menu for years, between DORA and two decades of security frameworks. And when three authorities feel compelled to restate the basics in 2026, the obvious question is how much of this exists in production and how much only in the slide decks that boards signed off on. The statement never asks that question out loud, which is a pity, because it is the most interesting one in the whole document.
Then comes the timeline, and this is where the exercise turns into involuntary satire. As Lead Overseer of critical ICT third-party providers under DORA, the ESAs will integrate AI-related risks into their oversight examination methodology and assess provider readiness, reflecting these themes in the scope of their 2027 activities. So the threat they describe operates continuously, adapts in real time and scales across systems, and the scrutiny of the providers everyone depends on will absorb that threat into its methodology next year. The statement itself is offered to entities and supervisors as a basis for supervisory dialogue, built on expectations that already existed. Dialogue. The attacker in their own scenario is an automated system probing infrastructure around the clock; the answer on the table is a conversation between a compliance department and a national authority.
There is also a quieter irony buried in the oversight section. The critical third-party providers that the ESAs will scrutinise for AI readiness include the very infrastructure on which the financial sector is concentrating its AI dependency. The concentration risk and the frontier AI risk are not two separate chapters; they are the same chapter, and the statement gets close to admitting it when it talks about shared infrastructure and supply chain compromise, before retreating into methodology and perimeters.
Look at the sequence as a whole: an ESRB warning in June, a Commission action plan in early July, a joint statement in late July, AI risks entering the oversight of critical providers in 2027, and in between, boards demonstrating commitment, which in corporate liturgy means one extra slide in the quarterly pack. Machine-speed offence, committee-speed governance. The gap between those two speeds is not a footnote to the systemic risk the ESAs describe. It is the systemic risk. And no supervisory dialogue, however cross-sectoral and risk-based, closes a gap that keeps widening while the paperwork travels through the perimeter of next year’s activities.
Raffaele Di Marzio
All my “insane” books on cybersecurity and governance are here 👇 https://cyberium.limited/bookshelf.html