AI-generated voices are becoming convincing enough to support high-value financial fraud, and cybersecurity professionals need to understand how these attacks work at the human and process level.
In a case disclosed in September 2026, fraudsters reportedly targeted Fideuram, the private banking arm of Intesa Sanpaolo. Reuters and Trending Topics reported that Fideuram’s then-chairman received a WhatsApp message appearing to come from Intesa Sanpaolo CEO Carlo Messina. A follow-up phone call used an AI-cloned voice of a senior lawyer known to him.
The operation resulted in overseas transfers totaling about €95 million. Around €53 million was later recovered, while approximately €36 million was still reported missing.
This case provides an important lesson for cybersecurity students and professionals: authentication can fail even when a person believes they have independently confirmed who they are speaking to.
What Is AI Voice Cloning?
Voice cloning uses artificial intelligence to generate speech that resembles a real person’s voice.
Modern tools can imitate characteristics such as tone, rhythm and pronunciation from recorded audio. Public figures and senior executives may have large amounts of voice material available online through interviews, conference recordings, webinars and corporate videos.
For an attacker, the objective does not need to be perfect imitation.
The synthetic voice only needs to be convincing enough in the context of an urgent conversation.
That is why voice cloning is particularly dangerous when combined with social engineering.
Why the Fideuram Attack Was Convincing
The reported attack did not rely on a single suspicious email.
It created a sequence.
First came an apparent message from a senior executive. Then another trusted person seemed to confirm the request by phone. Payment instructions were provided for the transaction.
To the recipient, this can feel like corroboration.
From a security perspective, however, verification is only independent when the verification channel is not controlled or supplied by the person making the request.
If an attacker provides the message, the telephone interaction and the payment details, those are not three independent proofs of identity. They are three components of one social-engineering operation.
Authentication and Authorization Are Different
This incident also illustrates an important cybersecurity concept.
Authentication asks: “Who is this?”
Authorization asks: “What is this person allowed to do?”
Even if someone believes an executive has been authenticated, a high-value payment should still be governed by authorization controls.
That might include dual approval, transaction thresholds, separation of duties, beneficiary verification and predefined escalation procedures.
A secure design assumes that identity checks can fail.
This is similar to defence in depth in technical systems. One failed control should not automatically lead to a catastrophic outcome.
Skills Cybersecurity Professionals Should Develop
Deepfake fraud makes several skills increasingly relevant.
Learn social-engineering analysis. Understand how urgency, authority, secrecy and familiarity influence decision-making.
Study identity and access management, because digital trust increasingly depends on strong authentication rather than human recognition.
Understand business processes. A cybersecurity professional who knows how payments, approvals and treasury workflows operate can identify weaknesses that a purely technical assessment may miss.
Develop incident-response skills too. In financial fraud, speed matters. Organizations may need to involve banks, law enforcement, legal teams and international partners quickly to freeze or recover funds.
Finally, learn how generative AI changes threat modelling. Ask not only whether a system can be hacked, but whether a trusted person, voice, video or instruction can be convincingly simulated.
Do Not Train People to “Spot the Fake”
Deepfake awareness is useful, but employees should not be expected to detect every synthetic voice.
As AI improves, that approach becomes unreliable.
The better strategy is to build processes that survive successful impersonation.
Use known contact details for independent callbacks. Require multiple approvals for sensitive transactions. Validate new beneficiaries separately. Treat unusual urgency as a reason to increase verification, not reduce it.
The Fideuram case shows why cybersecurity is no longer only about protecting machines.
It is about protecting decisions.
In the deepfake era, the safest question is not “Does this sound like my CEO?”
It is:“What independent evidence proves this request is genuine?”
To know more about Anand Shinde and his work in cybersecurity, awareness, and books:
https://anandshinde.com/
If your business needs cybersecurity services, AI security guidance, cloud security review, or protection against modern cyber threats, visit DevOM Consulting:
https://devomconsulting.com/
