A reported $108m voice-cloning fraud at Intesa Sanpaolo is the AI risk that is already here

Italy's Intesa Sanpaolo lost $108 million to an AI voice scam, according to the report circulating on this incident. Beyond the headline figure and the method — AI-generated voice used to impersonate someone the bank trusted — the publicly available detail is limited. We will not speculate about who was impersonated, how the transfer was authorised or whether funds were recovered, because that information is not in the material available to us.

What is worth saying is why this attack class keeps working. Banks have spent two decades hardening the digital perimeter: multi-factor authentication, transaction monitoring, device fingerprinting. The voice channel was never hardened to the same degree, because the voice itself was the authentication — you recognised the CFO, or the client, or the correspondent bank officer. Cheap, high-fidelity voice synthesis removes that assumption without touching a single line of the bank's code. The exploit is not in the systems; it is in the procedure that says a senior person's confirmed instruction is sufficient.

Our reading: this is the near-term face of AI risk, and it is unglamorous. Not a rogue superintelligence — a fraud playbook that used to require a skilled impersonator and now requires a few seconds of public audio, which means it scales. The defence is equally unglamorous and already known: out-of-band verification for large transfers, callback procedures to numbers on file rather than numbers offered by the caller, dual authorisation with mandatory delay, and staff training that treats urgency itself as the red flag. Banks that treat voice as an identity factor are running on an assumption that expired. The honest short-term reading is that fraud losses across the sector will get worse before controls catch up, and figures in this range will not stay exceptional. The honest long-term one is that this is a solvable operational problem — organisations fixed cheque fraud and card-present fraud with process, not prophecy — and solving it is a prerequisite for trusting AI with anything that matters.