An employee gets a call. They recognise the owner's voice, the tone, the way he phrases things. He asks for an urgent transfer to a new supplier, before the end of the day. The voice is fake, generated by software, but the money leaving the account is very real. This is the fraud that grew fastest over the past year, and it hits exactly the small firms that have no security department.
What actually changed
Voice cloning is not new technology. What is new is how cheap it has become, which means people with no programming knowledge at all are now using it.
Anthropic analysed 832 accounts banned for malicious activity between March 2025 and March 2026. The share of attackers rated medium risk or higher rose from 33 percent in the first half of that period to 56 percent in the second (Anthropic). The explanation matters more than the number: tasks that used to require years of technical experience are now done with a tool anyone can access.
On the damage side, the FBI introduced a separate category for fraud carried out with artificial intelligence for the first time, in its 2025 report: 22,364 complaints and roughly 893 million dollars in losses (SecureWorld, based on the FBI IC3 report for 2025). More than 30 million dollars of that comes from fake invoice and payment order scams, where a cloned voice was used to confirm an email by phone. In other words, exactly the scenario that could play out on a Tuesday morning in any company with two accountants and an owner who travels a lot.
It is already happening in Romania
The National Cyber Security Directorate has warned that deepfake technology is increasingly used in financial fraud, from cloning the voices of public figures to calls in which someone impersonates a manager or a colleague (DNSC, July 2026). The classic case, from 2019, involved a British energy company that lost 220,000 euros after a call using the imitated voice of the parent company's boss.
DNSC data for the first half of 2026 shows that 42 percent of reported fraud incidents in Romania were deceptive phone calls, in which the attacker can also fake the number displayed on screen (Realitatea). The same report carries two figures that explain why the trick works: cloned voice attacks rose by 442 percent in 2025, and in one test people correctly identified the fake voice in only 37.5 percent of cases.
The conclusion is uncomfortable: do not trust your ear. Neither you nor your accountant will tell the difference over the phone.
The rule that stops the attack, however good the voice is
There is one defence that works no matter how convincing the imitation, and it costs nothing: the callback.
Any new payment, any change of bank account, any urgent request that arrives by phone or email is confirmed by a call that you place, to the number you already have in the company's contact list. You do not call back the number that called you. You do not call the number in the email signature. You call the known number, the one from six months ago.
The attack rests on two ingredients: urgency and authority. A callback cancels both, because it moves the conversation to a channel the attacker does not control.
Three things to do this week, with no budget
1. Write the procedure on a single page. Who can approve payments, above what amount a second confirmation is required, how a change of bank account gets validated. Print it and put it next to the computer of whoever makes the payments.
2. Agree on a safe word. An ordinary word, known only to you and the two or three people who move money. If the call is genuine, they say it. If the person on the line hesitates, you have your answer.
3. Take the pressure off your staff. Tell them clearly, in writing, that nobody will be told off for delaying a payment by twenty minutes in order to verify it. Most of these frauds succeed because the employee was afraid of looking like they were ignoring the boss. As with the common mistakes small firms make when they start with AI, the problem is almost never the technology, it is the process around it.
Caution with your data: the part everyone forgets
The attacker's raw material is public information about your company: who the director is, how they speak, who handles invoices. A few seconds of voice from a clip on Facebook are enough for a decent clone.
The same care applies in the other direction, to what you send to an AI assistant. Do not paste contracts, client lists, personal identification numbers, bank account numbers or passwords into a chat window. If you need help with a real document, replace names with "Client A", delete the numeric codes and keep only the structure. For a company in Romania this is not excessive caution, it is a legal obligation: GDPR requires you to know where people's data ends up and who can read it. I covered the rules in the guide on what data you can and cannot give an AI assistant.
Since 2 August 2026, European rules require any artificially generated content to be labelled as such. Criminals ignore them, obviously, but you are obliged to follow them, a topic treated separately in the article about the obligation to announce that a chatbot is AI.
The concrete step for today
Open your company WhatsApp group or internal email and send a single message: "From today, any payment request or change of bank account received by phone is confirmed with a callback to the number in our contact list. No exceptions, including if the call comes from me."
It takes two minutes and it is the cheapest insurance policy you can buy this week.
Sources
- Anthropic, "What we learned mapping a year's worth of AI-enabled cyber threats" (3 June 2026): https://www.anthropic.com/news/AI-enabled-cyber-threats-mitre-attack
- SecureWorld, analysis of the FBI IC3 report for 2025 (9 April 2026): https://www.secureworld.io/industry-news/ai-enabled-fraud-topped-893m-fbi
- Financial Intelligence, DNSC recommendations against deepfake scams (11 July 2026): https://financialintelligence.ro/stiai-ca-nu-tot-ce-vezi-online-e-real-recomandarile-dnsc-pentru-a-te-proteja-inselatorii-prin-deepfake/
- Realitatea, DNSC data on online fraud in 2026 (5 August 2026): https://www.realitatea.net/stiri/stiinta-si-tehnologie/cum-te-feresti-de-fraudele-online-ce-arata-datele-din-2026-si-cum-te-protejezi-8lgpl7
Read next
Five common mistakes when a small firm starts with AI
GDPR and AI: what data you can and cannot give an AI assistant