Was it recorded?
Known-audio fingerprint checks ask whether the expected test content reached a recording. That is evidence about a recording path—not the identity of a caller.
VOICE RISK / DEEPFAKE DEFENSE
Recognize the voice.
Verify the authority.
Familiar audio is not sufficient authorization for a consequential instruction.
01 / THE THREAT MODEL
Trust the process, not just the sound.
A caller who sounds familiar can still be making an unauthorized request.
The FBI has warned that criminals use AI-generated audio to impersonate contacts, elicit payments and obtain access to accounts. That is reason to assess the risk—not evidence of a specific attack rate on trading desks.
Use independently established contact details or an authenticated channel to verify consequential instructions. Do not let the same unverified call supply both the request and its proof.
Read the FBI public warning ↗02 / THREE SEPARATE QUESTIONS
Recording. Synthesis. Authorization.
Was it synthesized?
Evaluate detection across codecs, noise, languages and unseen synthesis methods. Measure missed attacks and legitimate calls incorrectly flagged separately.
Was it authorized?
A real voice does not establish permission. Document the independent verification and approval needed before changing settlement details, access or payment instructions.
03 / CONTROL EXERCISES
Red-team the control. Not the customer.
A controlled exercise should end with evidence, not an unmanaged impersonation.
Design labeled synthetic-voice exercises around your firm’s controls. Before any evaluation, agree scope, source-audio rights, participant consent, approved channels and stop conditions.
- Written authorization and consenting participants
- Labeled test assets and a bounded exercise window
- No live transfers or unauthorized impersonation
- Human review and independent verification
- Separate false-alarm and missed-attack evidence
- Retained decisions, limitations and disposal rules
04 / FIELD NOTES
Go deeper on voice risk.
PERSPECTIVEVoice clones are coming for the trading desk
A measured approach to synthetic-voice risk. Read the article →
NEXT / DEFINE THE TEST
Start with the paths
you need to trust.
Bring your recording architecture. Let’s define the right test.