Prove the voice is live and present with an unpredictable challenge.
The threat
the attacker replays a recording (the victim's voice or pre-recorded answers, static cloned messages) instead of speaking live, to pass a voice check or a challenge.
Blind spotWhy classic frameworks miss it
a simple is-a-voice-here check does not tell live speech from a recording; the reuse risk is invisible if the audio input is not profiled (timestamps, stream).
MitigationProposed approach
audio stream fingerprinting (timestamps, RTP validation), anti-replay cache, unpredictable dynamic challenge (random phrase repetition), timing analysis.
The proposed control
a previously-seen stream is refused, the challenge cannot be replayed.
Expected evidence
a replayed recording is rejected before any action.
SourcesReferences and public research
MITRE ATLAS 2026.07AML.T0043 Craft Adversarial Data
Public researchPublic research sources: MITRE ATLAS 2026.07 (verified technique mapping), OWASP GenAI (model abuse categories), and the public risk-voicebot (aivansoul/risk-voicebot) template defining the 20 checkpoints. No client registry data: generic card, no rating, no verdict.