Rotterdam police arrested two suspects for bank helpdesk fraud in which AI-generated voices pressured victims into installing an app and having their bank card scanned via NFC, with around 170 victims and roughly €380,000 in losses. Protection requires controlling the whole chain: identity claim, time pressure, app installation, NFC action and rapid withdrawal, not just voice recognition.
The Rotterdam Unit of the police announced on 21 October 2026 that two suspects had been arrested. Our analysis is that, in this case, AI-generated audio served the telephone persuasion phase, while the subsequent cash withdrawals appear to have relied on NFC data from bank cards forwarded through the app.
What exactly did Rotterdam police establish about this form of fraud?
According to the publication by the Politie Eenheid Rotterdam, the criminals used dozens of AI-generated audio fragments with different voices, banks and scenarios to work on victims by telephone. Victims were made to install an app and hold their bank card against their phone, after which the NFC data — the contactless data on the card — was forwarded to an app used by a so-called pinner at a cash machine.
The police report around 170 victims and roughly €380,000 in losses over the investigation period. The case came about through cooperation between the police, banks and the Public Prosecution Service. The method therefore combines telephone manipulation, a technical app and simultaneous cash withdrawals.
Why do AI voices make bank helpdesk fraud more dangerous than before?
Our analysis is that AI lowers the threshold for building trust. In its explanation of how online criminals deceive victims, the Ministerie van Justitie en Veiligheid explains that with AI criminals can create more convincing fake messages, deepfakes and cloned voices, and that they use pressure, urgency and leaked personal data to win trust. In our analysis, a synthetic voice that sounds calm and professional can weaken the initial intuition that a conversation is not right.
In our assessment, the broader consequence is that recognisability as a line of defence weakens. In our analysis, the broader consequence is that recognisability as a line of defence weakens: AI-generated or cloned voices can reduce the reliability of vocal and conversational cues that people may otherwise use to assess a caller. The Fraudehelpdesk warns that false callers pose as trustworthy organisations, transfer calls to a fake supervisor and ask victims to install software through which attackers watch along remotely. In our analysis, cloned voices could reinforce the credibility of that escalation pattern.
The second shift is scale. The police describe dozens of pre-produced audio fragments with different voices, banks and scenarios. This shows how AI can support the preparation and variation of telephone fraud; the source does not, however, establish that this in itself explains how the group reached around 170 victims.
What does this mean for decision-makers, banks and fraud prevention?
Our analysis: because AI means the voice no longer functions as a reliable point of recognition, customers and staff lose their most important intuitive defence; therefore banks and organisations must now build their public information and internal training on behavioural rules that do not depend on voice recognition. Because the described cash withdrawals became possible after victims installed an app and had their bank card scanned via NFC, those actions are, in our analysis, key points where the chain can be interrupted; therefore unexpected, telephone-requested app installations and NFC actions should be technically blocked or immediately flagged as a fraud signal. Because the attack links several separate steps — telephone contact, new app registration, NFC use and cash withdrawal — our analysis is that detection on a single signal is likely to fall short; therefore fraud departments should consider monitoring precisely the anomalous combination of those events, such as a new app registration shortly followed by NFC traffic and a remote withdrawal. Because the Rotterdam case was only resolved through cooperation between the police, banks and the Public Prosecution Service, our analysis is that executives should now consider setting out which officer shares which relevant signal, and within what timeframe, with appropriate partners such as banks, police and the Public Prosecution Service, rather than improvising it per incident. Our analysis is that this chain resembles the vulnerability discussed in how errors spread through a linked chain: weaknesses at handovers could allow harm to propagate between steps.
For anyone considering voice AI in their own customer processes, this case also touches on governance around the provenance of voice data and voice AI. And because this is not a classic cyber incident but a human-focused attack with an AI component, it fits in a separate response playbook for AI incidents alongside the existing cyber runbook. More background is in our topic hub on AI security and threat management.
Which concrete behavioural rules protect customers and staff now?
According to the official fraud-prevention guidance cited here, customers should treat an unexpected telephone request to install an app or software as a fraud signal. In the Rotterdam case, an unexpected request to hold a bank card against a phone was also part of the fraudulent method. Every customer and staff member can test against that, regardless of how convincing the voice sounds. The Fraudehelpdesk advises ending the conversation, not downloading any programs and calling the bank yourself via a reliable number you have looked up yourself.
- When in doubt, end the conversation and call back yourself via a number on the bank card or the official site, not via a number the caller gives.
- Never install an app or utility at the request of an unexpected caller, not even from a so-called supervisor.
- Never hold your bank card against a phone at an unexpected caller's request; in the Rotterdam case, the app used the resulting NFC data for the fraud.
- Treat time pressure and urgency as a warning signal in themselves rather than as a reason to act.
- Report a suspicion immediately to the bank and the Fraudehelpdesk, so that signals quickly reach the chain.
Our analysis is that the impact of this kind of fraud depends less on the sophistication of the deepfake alone than on how quickly the chain from trust to cash withdrawal is interrupted. That is where the responsibility of executives and fraud departments lies: not only to assess the AI output, but also to interrupt or flag the steps that follow it.
Sources and references
Sources: The article relies on the publication by the Politie Eenheid Rotterdam, an explanation from the Ministerie van Justitie en Veiligheid and a warning from the Fraudehelpdesk.