Earlier this year, the ECB issued a supervisory letter to the CEOs of significant banks calling for proactive measures to address the escalating cybersecurity threats enabled by frontier AI models. Aishwarya Ramani, Senior Solutions Manager at Illumio, discusses AI resilience in the banking sector.
On July 7, 2026, Claudia Buch, Chair of the European Central Bank (ECB) Supervisory Board, issued a pivotal ‘Dear CEO’ letter (SSM-2026-0301), requesting major banks in the Eurozone submit an action plan addressing AI-enabled cybersecurity threats by October 31, 2026.
The message is not that banks face an entirely new class of risk. It is that frontier AI can expose and weaponise familiar weaknesses far faster than traditional remediation cycles can handle.
And speed is only half the problem. Vulnerability volumes are also increasing. The US National Vulnerability Database averaged around 80 new CVEs per day in 2023; by the first half of 2026, that had risen to 131. Security teams are being squeezed from both directions: more vulnerabilities to address and less time to do so.
Banks must still patch and modernise, but they cannot rely on remediation alone. They also need to limit how far an attacker can move when a vulnerability is exploited before a fix is available.
The collapse of the patch window
What really stands out in this letter is that frontier AI does not fundamentally change the underlying security issues banks face. Unpatched software, weak controls, exposed systems and ageing infrastructure are still the same risks they’ve always been.
What’s changed is the speed and scale at which those weaknesses can be exploited. Security teams have historically relied on a predictable cadence of vulnerability discovery, patch testing and scheduled maintenance windows. However, according to the Zero Day Clock, the time between a flaw’s public disclosure and active exploitation has shrunk from roughly ten months in 2021 to three hours in 2026.
Frontier AI models are driving that collapse by automating time-consuming elements of cyberattacks, such as identifying vulnerabilities, developing exploits and mapping potential attack paths. By its June 2026 update, Anthropic’s Project Glasswing had identified more than 10,000 high- and critical-severity zero-day vulnerabilities across participating critical infrastructure.
Vulnerabilities that were once manageable because defenders had time to respond become far more dangerous when attackers can operate at machine speed. Banks can no longer out-patch an automated adversary, nor should they treat every vulnerability as equally urgent. They need to identify which weaknesses create credible paths to critical systems, prioritise remediation according to business impact and accelerate patching where exposure is greatest.
Just as importantly, CISOs need the budget and board-level backing to tackle technology debt that has long sat on the backlog.
Building on DORA
That backlog won’t clear overnight, and banks don’t need to wait until it does to start containing risk. The EU Digital Operational Resilience Act (DORA) has already laid the foundations for digital operational resilience, including the ability to isolate affected systems and prevent an intrusion from spreading across critical operations. The ECB letter builds on that foundation for an environment in which AI is accelerating the speed of attack.
Crucially, the ECB is not asking banks to start again. The letter is explicit that the action plan should build on a bank’s existing cyber-risk strategy.
It also calls for “segmentation and, where feasible, microsegmentation, together with the adoption of Zero Trust principles, including continuous verification of users, devices, applications, APIs and service accounts”.
For banks, the direction is clear: assume a breach will happen, contain it quickly and limit the blast radius. That is the foundation of a Zero Trust architecture.
Shifting from detection to containment
Zero Trust works on the premise that no connection is trusted by default, regardless of where it originates. Every connection request must be continuously verified, and access must strictly adhere to the principle of least privilege.
Putting that into practice starts with visibility. Before a bank can decide which systems should be allowed to communicate, it first needs to see what’s actually happening across its network. Particular attention should be paid to east-west traffic. This is the traffic that flows between systems inside an environment and that attackers can exploit to move from an initial foothold towards critical assets.
By mapping those application dependencies, security teams can spot unnecessary connections and likely attack paths before they become a problem.
Visibility, however, is not the same as control. Illumio research found financial services to be the most confident sector globally at detecting unauthorised lateral movement, yet more than four in ten firms still admit they struggle to stop it once it’s underway. Seeing an attacker move is not the same as being able to stop them.
Microsegmentation is what closes that gap between detection and containment. Rather than trusting everything inside the network to communicate freely, it pushes the trust boundary down to the level of the individual workload, letting banks define exactly who can talk to whom and under what conditions. Everything else is blocked by default.
For banks facing faster AI-enabled attacks, that control provides something increasingly valuable: time. It gives security teams space to investigate, isolate affected systems and remediate vulnerabilities without allowing one compromise to become a much larger disruption.
Moving forward
For banking leaders, the priority now is not to try to eliminate every vulnerability. That is neither realistic nor the expectation. Instead, the focus should be on understanding which systems and services matter most, how an attacker could reach them and what controls are in place to stop a compromise from spreading.
In the age of frontier AI, the question is no longer simply, “Can we stop an attacker getting in?” It is also, “If they do get in, can we stop them getting any further?” Increasingly, that is where cyber-resilience will be won or lost.
