Regulation

# The ECB letter: look beyond the chatbot

The ECB asks banks to address AI-enabled cyber threats across their technology. A list of approved chatbots is only a small part of the answer.

2 September 2026 · 3 min read

Updated 28 October 2026

[← All articles](/blog)

[Start with the whole environment](#whole-environment)[Give each action an owner](#owners)[Include the AI work you actually run](#ai-work)[Use evidence for the question it answers](#evidence)[Sources and further reading](#sources)

## Start with the whole environment

The ECB's letter of 7 July 2026 asks the significant institutions it supervises for an action plan by 31 October 2026. It concerns AI-enabled cyber threats across their ICT environment, including how attackers use AI. It is not simply a request to govern the AI tools a bank buys.

That distinction matters. Removing a chatbot from a laptop does little about an attacker finding a weakness in an old internet-facing service. The attacker is unlikely to respect your approved-software list.

## Give each action an owner

Start with the systems, services and suppliers that support important work. Ask where a faster attacker could cause more harm and which existing weaknesses need attention. Include vulnerability management, detection, third-party risk and recovery.

For each gap, name an owner, the next action, a realistic date and a way to check the result. Separate measures already operating from work still to do. A delivery date is useful. Evidence that the measure works is useful too.

## Include the AI work you actually run

Internal AI tools can be part of this assessment. A coding agent may access repositories. A support assistant may read customer information. A tool-using agent may change a record. Identify the data, permissions and connected systems involved.

Choose a consequential workflow and test its controls. Can a request send sensitive information to an unapproved destination? Can an action proceed without the approval it needs? What happens when a required check is unavailable? Record the result and the routes the test did not cover.

## Use evidence for the question it answers

A patch record can support a patching claim. A recovery exercise can support a recovery claim. A sealed AI-action record can help a reviewer check what a connected control recorded. These are different kinds of evidence for different questions.

Athena can contribute data controls, action checks and verifiable records for connected AI workflows. That is a specific contribution to the evidence base. It does not replace an ICT-wide assessment or produce the whole action plan.

Read the original letter with your security, technology and supervisory teams. Confirm what applies to your institution and what has changed since the letter. Keep the wider plan and the narrower product contribution clearly separate.

## Sources and further reading

Originally published 2 September 2026. Rewritten on 28 October 2026.

[ECB letter, 7 July 2026](https://www.bankingsupervision.europa.eu/press/letterstobanks/shared/pdf/2026/ssm.2026_letter_on_AI_enabled_cybersecurity_threats.en.pdf)

Keep exploring

[Explore Athena](/athena)[Trust and verification](/trust)[All articles](/blog)
