Fideuram's €95 million deepfake: when the confirmation call is fake too

It started with a WhatsApp message that appeared to come from the CEO of Italy's biggest bank. Then came a phone call from a senior partner at a prominent law firm, confirming everything. Both were fake. By the time anyone noticed, Fideuram, the private banking arm of Intesa Sanpaolo, had sent about €95 million to overseas accounts, according to Reuters reporting published last week. Authorities have recovered roughly €53 million. Around €36 million is still missing, laundered through a network of accounts and converted into crypto.

The headline number is striking. The mechanics are what fraud teams should study, because this attack did not beat a weak control. It beat the control most organisations treat as their safety net: independent confirmation.

How the Fideuram scam worked

According to the sources cited by Reuters, the scheme unfolded in February in three moves:

1. The request. Fideuram's then-chairman received a WhatsApp message that appeared to come from Intesa Sanpaolo CEO Carlo Messina, asking for urgent help with an overseas transaction.
2. The corroboration. A follow-up phone call appeared to come from a senior partner at a well-known law firm, confirming the request. The lawyer's voice was cloned with AI.
3. The execution. Believing the instructions were legitimate, the chairman asked the finance department to make a series of transfers, mainly to accounts in China and Hong Kong.

No Fideuram executives are under investigation. Milan prosecutors are reportedly investigating a foreign national living outside Europe. The takeaway is uncomfortable: a senior, experienced banker saw a request corroborated by a second trusted source and still got it wrong.

When the confirmation is the attack

Most payment-fraud playbooks share one rule: if a request arrives through an unusual channel, confirm it through another one. Fraudsters know that rule too. In the Fideuram case, the second channel was part of the script. The voice call didn't expose the fake WhatsApp message. It validated it.

This is the shift security leaders need to internalise. Classic business email compromise relied on a single forged message and hoped nobody picked up the phone. Voice cloning lets attackers fabricate the confirmation step as well, borrowing the authority of a third party, such as a lawyer, auditor or regulator, who the target has no reason to doubt and no habit of verifying.

It isn't the first time in Italy. Last year, fraudsters mimicked the voice of an Italian government minister to persuade businessman Massimo Moratti to transfer nearly €1 million. The pattern is consistent: impersonate someone with authority, add urgency, and supply the "proof" yourself.

Why voice trust is eroding across banking

Fideuram is one high-profile data point in a much broader trend. A recent analysis in the ABA Banking Journal highlights how quickly the voice channel is being weaponised:

• Voice fraud in banking rose by roughly 30% in 2025.
• Synthetic voice attacks against financial institutions increased by a factor of about 20 in a single year.
• Roughly one in 750 banking calls was flagged as potentially fraudulent in the same period.
• Many detection tools that perform well in the lab lose accuracy on real phone audio, where noise, compression and channel distortion are the norm.

The same analysis notes that banks are moving toward approaches that combine voice analysis with behavioural and transactional context, rather than relying on any single signal.

What fraud and treasury teams should change now

The Fideuram case points to specific gaps that apply well beyond private banking.

Treat corroboration as untrusted input

An inbound confirming call proves nothing on its own. Verification should go out to contact details held in your own records, never to numbers, links or contacts supplied as part of the request. If a "lawyer" calls you, you call the firm's switchboard back.

Remove single points of authority

The chairman's instruction was enough to move money. High-value or unusual transfers should require dual approval from people who verify independently, regardless of seniority. Attackers target the top precisely because nobody questions it.

Verify the voice, not just the story

Human judgment cannot reliably tell a cloned voice from a real one. Real-time audio deepfake detection on executive, treasury and high-risk call flows gives staff an objective signal at the moment they're being pressured to act.

Bind the voice to a verified identity

Detection answers "is this synthetic?" Identity assurance answers "is this the person they claim to be?" Combining both, for example by checking that a caller's voice is consistent with their verified face on video, raises the bar well beyond what a cloned voice alone can clear.

The new standard: prove the person, not the channel

The Fideuram fraud shows that channels, caller IDs and even a second confirming voice can all be manufactured. What attackers cannot easily fake at scale is a biometrically consistent, real-time human presence that holds up to machine analysis. Banks that keep relying on "it sounded right" will keep paying for it.

Corsound AI's Deepfake Detect analyses audio and video in real time to flag synthetic voices before money moves, and pairs with Voice-to-Face AI to confirm that the voice on the line belongs to the person on screen. See how Deepfake Detect protects high-value calls.