The wire-transfer scenario, set in 2026
Deepfake-CEO wire fraud is real — the intervention point is 90 seconds before the transfer, when someone runs the clip through a check.
A clip arrives in your finance team’s inbox. It looks like a video call recording with the CEO. The CEO is asking for a $4.2 million wire to close an acquisition before the EU window shuts. Voice matches. Face matches. The urgency matches the CEO’s actual schedule. Send the wire today, the CEO says, we’ll do the paperwork tomorrow.
This is the playbook that settled $25 million out of an Arup engineering office in Hong Kong in 2024. The synthetic-media tooling has gotten cheaper and faster every quarter since. The intervention point isn’t the wire-approval workflow. It’s ninety seconds before that. Someone on the finance team drops the file into a deepfake check and waits for a verdict.
That is what this product is.
What we actually look at
Three signals — face artifacts, voice splices, and metadata — feed one four-band verdict, each band tied to a clear action.
Three pipelines run in parallel on every upload: face-artifact analysis, voice-splice detection, and metadata-and-provenance checks. Each looks for the tells its layer leaves behind. See the trust ledger for the forensic detail and our published benchmarks.
The three feed a verdict in one of four bands, each tied to an action. Real means trust the source. Probably real means trust but verify out of band. Probably fake means don’t act on this. Fake means escalate.
What we are honest about
We are not a courtroom expert witness. We will sometimes be late against a generator that ships the same week we run the verdict. The model arms race is real. So we publish accuracy benchmarks against current commodity generators in the trust ledger rather than hand-wave at it. The verdict is a 90-second editorial sanity check. It catches the large majority of current deepfake content, with very low false-positive rates against the probably-fake and fake bands. For litigation, we refer you to qualified forensic vendors with the right chain-of-custody discipline.
Personalisation under Executive
On Executive, verdicts run against your enrolled voiceprint — answering not just “is this fake?” but “is this really you?”
A generic deepfake verdict says “this voice was probably spliced.” When the speaker is an enrolled customer, an Executive verdict adds “and it does not match your enrolled voiceprint.” That second answer is the one that matters. The question is whether your CEO actually made the call. See Executive Identity Defense and Voiceprint Protection for the enrolment and consent detail.