Europe's bank supervisor wants an AI-cyberattack plan from every bank by October 31: the template any board should borrow

In July the European Central Bank told every bank it directly supervises to write a plan against AI-powered cyberattacks, due this Friday. American Banker ties the urgency to Claude Mythos, the Anthropic model kept under tight restrictions after it found thousands of previously unknown vulnerabilities in major operating systems and browsers. Four steps to write that plan even if your company isn't a bank.

On July 7, Claudia Buch, chair of the European Central Bank's Supervisory Board, wrote to the chief executives of every bank the ECB directly supervises, asking for a written plan against AI-powered cyberattacks. The deadline she set, confirmed by American Banker under the headline "ECB tells banks: Fix AI cyber gaps by Oct. 31," falls this Friday, October 31. It is not a proposal or a generic recommendation: the letter, referenced SSM-2026-0301, requires each institution to describe concrete measures, name an internal owner, commit budget and staff, and set execution deadlines, then hand the plan to its Joint Supervisory Team. The recipients include the euro-area subsidiaries of JPMorgan Chase, Goldman Sachs, Citi and Morgan Stanley.

Per KPMG's rundown of the letter, the ECB asks banks to act on six fronts: prioritize the most vulnerable areas, speed up vulnerability and patch management, deploy AI-assisted detection and response, strengthen governance and third-party risk, apply defense-in-depth built on zero-trust principles, and prepare crisis management and recovery. KPMG maps those six areas onto the DORA articles already in force: the letter treats AI as a multiplier of speed and scale on risks banks already had to manage, not as a new rulebook. It also does not threaten a specific fine for missing the deadline: a cyber-regulation lawyer quoted by American Banker said the ECB may allow leeway for a good reason, but if unsatisfied could escalate to a formal order to comply, financial penalties, or limits on new business.

The trigger: a model Anthropic keeps tightly restricted

American Banker ties the urgency of the letter to a specific industry event: earlier this year Anthropic said that a model it distributes only under tight restrictions, Claude Mythos, had found thousands of previously unknown vulnerabilities across major operating systems and web browsers, and the company has kept access to it tightly restricted. The ECB's letter does not name the model. The detail matters less for the model itself than for what it signals: a capability deliberately held back today could, in a future generation or in less cautious hands, become an attack tool available to anyone.

Why it matters even if your company isn't a bank

The ECB can only impose requirements on the institutions it supervises, but the reasoning behind its letter isn't specific to banking: if an AI model can find and weaponize software flaws faster than a company can patch them, the patch cycle stops being the defense it used to be. That problem reaches any company with internet-facing systems, outside vendors and open-source components, which today is practically every company.

Four steps to write that plan before anyone asks for it

  • Treat speed as the new variable, not the attack surface. The ECB's concern isn't that vulnerabilities exist, everyone already knew that, it's that an AI-assisted attacker can find them and turn them into a working exploit in far less time than a human team needs to patch them. Measure your current patch cycle against that assumption, not against the one from two years ago.
  • Put it in writing with a name, a budget and a date, even if no one is asking yet. The template the ECB demands, concrete measures, a named owner, assigned money and staff, and a deadline, works just as well as a voluntary internal plan as it does as a supervisory requirement.
  • Check what isn't yours first. Outside vendors, third-party software and open-source components sit alongside a bank's own systems on the ECB's list; for most companies, that is exactly where real visibility is lowest.
  • Rehearse the response before you need it. A crisis plan no one has tested is, in practice, no plan at all; one annual drill with the people who would actually have to act is worth more than the document describing it.

On our side, when we help a company look squarely at its AI exposure, we start with exactly the question the ECB has now made formal: who would be accountable, with what budget and on what timeline, if this week's security flaw is found by a model before it's found by a person.

Sources

← Back to the blog NHAI.AI — AI news

Put AI to work with expert supervision.

Agents from $79/month, or a free Express Diagnostic to see where AI fits your business.

See agents & pricing