ECB AI Cybersecurity Action Plan: What Banks Must Submit by October 31

How to build a supervisor-ready AI cybersecurity action plan before the October 31, 2026 deadline

On July 7, 2026, the European Central Bank sent a letter (SSM-2026-0301) to the CEOs of every significant institution under its direct supervision. By October 31, 2026, each of them must submit a concrete action plan to its Joint Supervisory Team (JST) addressing AI-enabled cyber threats.

The ECB’s premise is blunt: AI models can now discover software vulnerabilities and generate working exploits at a speed that makes traditional patch cycles obsolete. Attack timelines that used to run in weeks now run in hours. If that’s the threat model, a defense strategy built on periodic assessment is outdated by design.

This post covers the essentials: what the ECB is asking for, what separates a plan that survives scrutiny from one that invites challenge, and how to build the evidence behind it.

What the ECB is actually asking for

The letter identifies six priority areas: protecting attack surfaces, accelerating vulnerability and patch management, strengthening AI-enabled monitoring and defense, reinforcing governance and supply chain assurance, advancing defense-in-depth, and improving operational resilience.

Read together, they add up to something more demanding than a planning exercise. A JST reviewing hundreds of action plans will separate them on one axis: which plans are built on assumptions, and which are built on real exploitability and evidence. And you won’t be judged in isolation: the ECB will run a horizontal analysis across every submitted plan, so yours is read comparatively, against your peers.

That’s the core argument: this is a proof mandate disguised as a planning mandate. A vulnerability scan tells you what might theoretically be exploitable. Security validation tells you what actually is. Pentera proves it the way an attacker would: safely exploiting your live environment, continuously and under deterministic control, so every exposure you report is backed by a real, reproducible attack rather than a severity score. That is what turns an action plan from a list of intentions into evidence a supervisor can rely on.

Answering machine-speed attackers with machine-speed validation

The letter is explicit: a prudent posture “assumes that perimeter defences will be breached.” An assume-breach stance can’t be asserted on paper; it has to be tested. And when the time from disclosure to weaponization approaches zero, the question “are we exploitable right now?” can’t be answered by an annual pentest or a vulnerability assessment sorted by CVSS. It has to be answered continuously, by safely exploiting your own environment the way an attacker would. That is what automated security validation does.

The ECB is also clear on how AI belongs in that defense: AI-enabled tools are acceptable only with “governance, validation and human oversight.” Deterministic, repeatable validation meets that bar. Autonomous tooling that improvises a different attack on every run does not.

Continuous, exploit-based testing changes what an institution can put in front of its supervisor:

• A real map of the attack surface, internal, external, and cloud, built from live adversary emulation, not asset inventory assumptions. Pentera Surface covers the external view; Pentera Core validates the internal network; Pentera Cloud extends it to cloud environments.
• Risk prioritization grounded in proven exploitability. Validated attack paths to critical assets, not theoretical severity scores. Pentera Resolve routes proven exposures into remediation workflows and re-validates once fixes land.
• Evidence that defense-in-depth holds. Testing segmentation and lateral movement under real attack conditions, the difference between an architecture diagram and a defensible claim.
• The identity attack path. Weak, reused, and leaked credentials harvested and replayed to escalate privilege and move toward critical systems, the way real attackers get in. The letter singles out least privilege, MFA, and access controls, and identity is where most breaches actually start.
• Resilience tested against the scenario that matters. A ransomware readiness assessment run safely in production gives crisis-management planning a live-fire baseline.
• Threat-informed validation. Pentera’s strategic partnership with Recorded Future pulls live threat intelligence, leaked credentials, weaponized vulnerabilities, and active-actor activity, straight into the platform, then safely exploits those exact threats against your own environment. You get proof of whether what attackers are using right now actually reaches anything in your estate, not just a dashboard telling you the exposure exists.

This is continuous validation doing what point-in-time testing can’t: producing dated, repeatable, auditable evidence a regulator can rely on.

One evidence program, two regulations

Significant institutions don’t face the ECB letter in isolation. The ECB is explicit that DORA’s requirements “remain highly relevant and valid”: this letter does not replace DORA; it raises the bar on the evidence DORA already demands. That includes Threat-Led Penetration Testing (TLPT, Articles 26–27) against live production systems. Continuous validation keeps you TLPT-ready between those mandated exercises and produces the same evidence your ECB plan needs. The same program directly supports DORA compliance and ICT risk management.

What a supervisor-ready plan actually contains

Four evidence artifacts separate a plan that holds up from one that invites challenge:

• A validated baseline: which attack paths were actually exploitable, proven by safe exploitation in production, across internal, external, and cloud surfaces as of a dated run.
• A prioritization rationale regulators can’t argue with: remediation ordered by proven exploitability and business impact, not CVSS scores.
• Re-validation proof: for every remediated finding, evidence the fix was tested against the original attack and the path is closed.
• Delta reporting: posture change across runs, showing exposure trending down over time.

Only continuous, exploit-based validation produces all four. Institutions already running it can assemble the evidence today; those that aren’t have roughly 60 days to establish a baseline before the deadline.

Where this fits in your exposure management program

For institutions running a CTEM program, the ECB letter effectively regulates pillar four: validation. Exposure management that stops at discovery and prioritization, without adversarial exposure validation, produces exactly the assumption-based plans the ECB is warning against. That is the gap the Pentera platform closes, across every attack surface.

On third-party risk specifically, Pentera proves whether the third-party and open-source software running inside your environment is actually exploitable, the complement to vendor-rating tools that score a supplier’s external posture.

The bottom line

The ECB has written the business case for security validation into a supervisory mandate. Its core claim, that AI attackers move faster than assumption-based defense, is the same claim Pentera has built its platform on. October 31 is the submission date, not the finish line: the JST monitors progress and compares plans across peers, so yours has to keep producing evidence after you submit it. That is what continuous validation does that a point-in-time exercise cannot. You cannot out-plan a machine-speed attacker. You can out-prove one.

You don’t have to build this alone. Sector 11 can augment your team through the submission window, running the validation, assembling the four evidence artifacts, and packaging them into a JST-ready submission, so a stretched security function can still produce a defensible, proof-backed plan by October 31.

Talk to us about establishing your validated baseline before October 31.

Pentera named a Customers’ Choice in the 2026 Gartner® Peer Insights™ Voice of the Customer* for Adversarial Exposure Validation

The recognition comes straight from the security teams who use Pentera — verified customer reviews, not vendor claims. That’s the same principle behind this whole guide: proof over assertion. Read the report →

*GARTNER is a registered trademark and service mark, and PEER INSIGHTS is a trademark of Gartner, Inc. and/or its affiliates. Gartner Peer Insights content consists of the opinions of individual end users based on their own experiences and does not represent the views of Gartner or its affiliates.