Back

Deepfake voices in the approval chain

The three-second voice clone against the person most likely to authorise a wire. Why call-back to a fixed number is the only control that consistently holds.

From Anieres

Three seconds of someone's voice is enough to build a synthetic version that will convince a colleague on a routine call. The person most likely to authorise a wire from a family office is also the person whose voice is most reliably recorded in public.

A family office whose payment approvals rely on a verbal confirmation from a principal, particularly when the principal travels and the confirmation arrives from an unfamiliar number, is exposed to a class of loss the payment policy usually misses. The attack is mature and has a defined operational tempo.

Voice recognition on its own does not hold, because a synthetic voice built from public audio will pass it. Contextual questions do not hold at any useful rate. The control that consistently holds is a call-back to a fixed, pre-agreed number on a channel independent of the one the request came in on, before the payment is released, with no exception for urgency and no exception for the seniority of the caller.

Define the fixed call-back number now and write into the payment policy that no other control substitutes for it. A control designed after the incident is a control that will be argued with.

Written alongside work at Anieres: exposure mapping, cross-reference, and standing-report systems for private clients.