Topic 5 · Deep Dive

5.6 Multi-Modal AI Threat Vectors

Voice Cloning Mechanics

Modern zero-shot text-to-speech systems need only seconds of a target's speech, and those seconds are usually already public: voicemail greetings, conference keynotes, podcasts, and social video. The same class of models also runs in real time as voice conversion over live calls, which powers the vishing playbooks you should rehearse: the IT helpdesk offering a "courtesy password reset," the bank's "fraud team" calling ahead of a fabricated incident, and the classic urgent executive wire request. Synthetic voice leaves detectable artifacts: flattened prosody and unnatural pause placement, missing breath and mouth noise between phrases, spectral smoothing that sounds faintly metallic or over-clean, room tone that shifts between words, and suspiciously wideband, studio-grade audio arriving over a phone channel.

Synthetic Video in Meetings

Video impersonation in the wild is not a rendered file; it is a live puppet. Real-time face-swap and reenactment models are injected into a call as a virtual camera feed, so every frame is generated at call time from the attacker's own face and voice. The pretext is already standardized: a "bad connection" excuse that justifies low frame rates, and a camera-on request aimed at the victim, not the impersonator. What to watch for: lip-sync drift that worsens on sibilants, abnormal blink rates, edge tearing or blur at the jawline and hairline during head turns, texture jitter under compression, and lighting on the face that contradicts the background.

The Trust Heuristic Is the Bypass

What these vectors defeat is not a technical control but a human heuristic: voice plus face functioned as two live identity factors only while impersonation was expensive. Generative tooling collapsed that cost, so the heuristic is now stale -- a familiar voice is no longer proof of identity, and a camera feed is no longer proof of presence. Detection-first defenses also misfire, because the decision happens in seconds on a live call. Replace the heuristic with procedure instead: never authorize a payment, a password reset, or a disclosure based on inbound media alone; call the person back on a number pulled from the corporate directory, never one supplied in the call itself; keep a shared challenge phrase for sensitive confirmations; impose mandatory verification delays on financial changes; and re-authenticate over a known channel, such as a message inside a verified workspace.

Architecture Diagram

multi-modal impersonation lanes cloned voice vishing call synthetic video meeting impersonation human trust bypass deepfake counter: callback on official number + challenge phrase
Voice and video lanes differ in delivery but converge on the same defect: a trust heuristic priced for a pre-deepfake world.

Key Takeaways

« Back to Topic 5« 5.5