What happened in the Fideuram fraud case?

Fraudsters sent the bank's then-chair a WhatsApp message impersonating Intesa Sanpaolo's CEO, then followed with a call in an AI-cloned voice of a senior lawyer, persuading him to order transfers of 95 million euros to accounts in China and Hong Kong. About 53 million euros was recovered.

How does the case compare to the Arup deepfake fraud?

Arup's 2024 loss, $25 million, required a full deepfaked video meeting; Fideuram's required only a spoofed messaging identity and one cloned voice call, showing the lighter attack chain scales to larger losses when aimed at senior authority.

How do attackers clone executive voices?

From public audio: earnings calls, conference panels, podcasts, and video scraped from platforms like LinkedIn supply the training data, which is why executive media exposure now doubles as the adversary's enrollment corpus.

What controls stop CEO impersonation fraud?

Ceremony, not intuition: high-value instructions verified through enrolled channels with callbacks to directory numbers, liveness-anchored identity proof at approval, dual authorization over independent channels, and approvals dynamically bound to the exact amount and payee.

How big is AI-driven fraud expected to get?

Deloitte's cited projection puts generative-AI-enabled US fraud losses at $40 billion by 2027, with current incident data showing the US as the most-targeted country and executive impersonation among the fastest-growing categories.

Back to News

The Deep Brief · Oct 16, 2026 · 4 min read

A WhatsApp message and a cloned voice cost a bank $108M

Fideuram lost 95 million euros to AI messaging fraud: a fake CEO WhatsApp plus a cloned lawyer's voice moved funds to Asia, with only 53 million recovered.

Rosalie Chirip

Senior Editor at deepidv

A smartphone in the middle of a voice call with the red end call button lit, the kind of call a cloned executive voice can place.

The deepfake fraud era has its second nine-figure case study, and this one needed no video call. Fideuram, the private banking arm of Intesa Sanpaolo, lost 95 million euros, about $108 million, to an AI-assisted impersonation chain first reported by Reuters: a WhatsApp message to then-chair Paolo Molesini, appearing to come from Intesa Sanpaolo chief executive Carlo Messina, requesting urgent help with an overseas transaction, followed by a phone call in an AI-cloned voice imitating a senior lawyer at an associated firm. The combination was enough. Molesini instructed his finance department to arrange transfers to foreign accounts, mainly in China and Hong Kong; authorities have recovered roughly 53 million euros, the remainder having moved overseas and into cryptocurrency, and Milan prosecutors have a foreign national under investigation for computer fraud.

The case reads as the messaging-fraud counterpart to 2024's defining incident, the Arup case, where a Hong Kong employee wired $25 million after a Zoom call with deepfaked executives. Together they bracket the attack surface: Arup proved the synthetic video meeting, Fideuram proves the lighter chain, a spoofed text identity plus one cloned voice, works at four times the price. The operational lesson is harsher in the second case, because nothing in it was technically sophisticated: no compromised systems, no forged documents, just impersonated authority over unverified channels, aimed at the top of the hierarchy where deference does the attacker's work.

The numbers behind the trend

Italian detection firm identifAI frames the case inside the pattern this publication's deepfake coverage has tracked all year, and its dataset adds texture: the US is the most-targeted country at 34 percent of incidents, X carries 42.1 percent of platform-linked cases with LinkedIn second at 14.5 percent, and, pointedly, attackers scrape platforms like LinkedIn for executive audio and video to train impersonation models, meaning every earnings call and conference panel is now enrollment data for the adversary. Deloitte's projection, cited in the report, puts generative-AI-driven US fraud losses at $40 billion by 2027, and the recommended defenses align with the stack this publication keeps mapping: cryptographic media authentication, continuous behavioral signals, and verification that moves beyond legacy checks, with the FTC's Impersonation Rule supplying the US regulatory hook.

The ceremony that was missing

Strip the AI and the Fideuram loss is a ceremony failure: a nine-figure payment authority executed on the strength of an unauthenticated message and an unverified call. The defense is correspondingly procedural, and it is the one treasury teams have resisted as friction for years: high-value instructions bind to verified humans over enrolled channels, full stop. A callback to a directory number, not the number that called; identity proof at the approval moment, liveness-anchored, because cloned voices pass ears and fail structural checks; dual authorization with independent channels for amounts that matter; and dynamic binding of the approval to the specific amount and destination, so a redirected payee voids the authorization. Every element existed before generative AI; generative AI merely removed the last excuse that the voice on the phone was proof of anything. The banks that internalize Fideuram will treat executive identity as a verification problem with the same rigor they apply to customers, and the attackers' own economics, $108 million for a text and a voice model, guarantee the next attempt is already in someone's inbox.

Fideuram Impersonation Fraud FAQ

What happened in the Fideuram fraud case?: Fraudsters sent the bank's then-chair a WhatsApp message impersonating Intesa Sanpaolo's CEO, then followed with a call in an AI-cloned voice of a senior lawyer, persuading him to order transfers of 95 million euros to accounts in China and Hong Kong. About 53 million euros was recovered.
How does the case compare to the Arup deepfake fraud?: Arup's 2024 loss, $25 million, required a full deepfaked video meeting; Fideuram's required only a spoofed messaging identity and one cloned voice call, showing the lighter attack chain scales to larger losses when aimed at senior authority.
How do attackers clone executive voices?: From public audio: earnings calls, conference panels, podcasts, and video scraped from platforms like LinkedIn supply the training data, which is why executive media exposure now doubles as the adversary's enrollment corpus.
What controls stop CEO impersonation fraud?: Ceremony, not intuition: high-value instructions verified through enrolled channels with callbacks to directory numbers, liveness-anchored identity proof at approval, dual authorization over independent channels, and approvals dynamically bound to the exact amount and payee.
How big is AI-driven fraud expected to get?: Deloitte's cited projection puts generative-AI-enabled US fraud losses at $40 billion by 2027, with current incident data showing the US as the most-targeted country and executive impersonation among the fastest-growing categories.