the video call deepfake alarm is coming, but attackers aren't waiting for it

It took one video call and fifteen wire transfers for a finance employee at a global engineering firm to lose $25.6 million. Every other person on that call, including the company's own CFO, was an AI-generated deepfake — synchronized facial movements, cloned voices, and all. It remains the largest publicly known case of live deepfake fraud, and it's the backdrop against which researchers in Germany just unveiled a new kind of defense: a system that flags deepfakes while a video call is still happening.

A live warning system for fake meetings

Engineers at Fraunhofer SIT and Fraunhofer Heilbronn have built a prototype that analyzes both audio and video streams during a live meeting and generates a running probability score that a participant is synthetic. When the score crosses a threshold, the system prompts attendees to verify identity through a separate channel or additional questions. Crucially, it runs locally rather than sending sensitive footage to an external server — a meaningful design choice for banks, law firms, and government agencies wary of routing board-level conversations through third-party cloud infrastructure.

The research team says it wants to work with videoconferencing providers to eventually build the capability into platforms like Teams or Zoom. That's a genuinely useful direction for the industry. It's also, for now, a lab prototype — not something a bank's fraud team can deploy on Monday morning.

Why a prototype isn't the same as protection

Compressed video streams lose the pixel-level artifacts that older deepfake detectors relied on, which is exactly why real-time detection is hard and why this kind of research matters. But the gap between a promising prototype and a shipped, enterprise-grade product is typically measured in years, not months. Meanwhile, deepfake fraud attempts have climbed 2,137% over the past three years, and 62% of organizations reported experiencing a deepfake attack in the prior 12 months, according to a recent Gartner survey. Combined voice and facial deepfake checks are projected to more than double globally by 2028. The threat is scaling faster than academic research can productize a fix.

What this means for fraud and security teams today

Waiting for a future Teams or Zoom update to solve live-meeting fraud is not a defensible strategy for organizations handling wire transfers, customer onboarding, or sensitive HR conversations right now. A few practical steps close most of the gap immediately:

• Verify high-value requests out-of-band. Any instruction to move money or change account details that arrives via video or voice should be confirmed through a second, independent channel before execution.
• Deploy detection at the point of the call, not after the loss. Real-time audio and video analysis needs to run during the meeting, not as a forensic exercise once funds are already gone.
• Treat voice and face as one signal, not two. Attackers increasingly clone both simultaneously, so verification that only checks one modality misses the other.
• Establish escalation protocols for executives. Pre-agreed verification phrases or callback procedures for C-suite requests remove the social pressure that makes these attacks work.

The defense researchers are racing toward already exists

What Fraunhofer's team is prototyping in the lab — continuous, real-time analysis of both audio and video to catch synthetic participants mid-call — is the same problem Corsound AI's Deepfake Detect was built to solve for enterprises today. Rather than waiting for a future platform update, banks, telecoms, and HR teams can deploy real-time audio and video deepfake detection now, at the moment a call happens, before a fraudulent instruction ever reaches a wire desk. Research prototypes point toward where the industry is headed; production-grade detection is what protects the meeting happening this afternoon.

See how real-time deepfake detection works today at corsound.ai/deepfake-detect.

Photo: Christina Morillo / Pexels