A consented, generated adversarial voice-call corpus

BREACHER.AIManage supplier listing

A consented, generated adversarial voice-call corpus. Every campaign produces paired synthetic-and-real telephone audio: a cloned voice with signed likeness authorisation holds a live two-way conversation with an employee, answers questions, and independently handles inbound callbacks, running through to the help-desk password-reset or finance-approval step. Each session therefore yields full-call audio with speaker turns, the pretext and script used, the verification challenge issued, which specific control step was skipped or satisfied, the final process outcome, and — critically for deepfake detection — ground-truth labels on exactly which audio segments are synthetic. Published evidence of what the corpus shows: 63% of targets could not distinguish synthetic voice or video from genuine during the call, and 78% of assessed organisations were rated highly vulnerable when measured as process failure rather than individual error