What is driving deepfake detection into live conversations?

AI-generated audio and video paired with social engineering now target calls, meetings, and contact centers directly, and vendors have concluded human participants cannot reliably detect manipulation aimed at their psychology, so detection is moving into the channel itself.

What did Netarx launch?

An AI Defense API built on an orchestration layer of 40 specialized models cross-validating signals in real time, spanning device fingerprints, EXIF data, compression signatures, audio micro-artifacts, and video tells such as blink frequency and lighting inconsistencies.

What is Pindrop BotStopper?

A standalone detector that identifies AI agents and automated callers in about two seconds, backed by a registry of more than 5,000 AI voices. Pindrop positions it as Know Your Agent for voice: identify the agent and its principal rather than block all machine callers.

Are the vendor accuracy claims verified?

No. Detection rates, the two-second identification figure, and the healthcare case study with its 94.3 percent reduction are company statements, and registry-based approaches only recognize enrolled voices, so independent testing remains essential.

How should firms combine conversation-layer detection with identity verification?

Use synthetic-media detection as one evidence stream and fuse it with identity, device, and liveness signals, so high-risk actions requested mid-call route to verified, human-bound approval ceremonies rather than resting on a single detector's verdict.

Back to News

The Deep Brief · Oct 23, 2026 · 4 min read

Deepfake defense moves into the live conversation

Netarx, Onecom and Pindrop shipped real-time deepfake and bot detection for calls and meetings this month, as fraud attacks target the conversation itself.

Rosalie Chirip

Senior Editor at deepidv

A laptop showing a video meeting grid of remote participants beside a green mug, the setting where live deepfake checks now run.

The identity industry has decided where the next fight happens: inside the call. A cluster of October launches, rounded up by Biometric Update, moves deepfake detection out of onboarding flows and into live conversations, meetings, and contact centers, on the shared diagnosis that AI-generated audio and video plus social engineering have made the conversation itself the attack surface. The thesis was put plainly in a SecurityWeek analysis the roundup cites: human defenders should not and cannot be expected to detect trickery designed to manipulate their psychology.

Netarx, the Michigan-based entrant, launched an AI Defense API built on an orchestration layer running 40 specialized models cross-validating thousands of signals across modalities in real time. The checks read like a forensics syllabus, device fingerprints, EXIF data, compression signatures, and location mismatches on the hardware side, micro-artifacts in audio, and micro-expressions, blink frequency, lighting inconsistencies, and hairline warping in video. Founder Sandy Kronenberg's framing captures the shift: "for decades we protected the network. Now the attack surface is the conversation itself." Onecom's next-generation Vox platform brings the same instinct to business telephony, pairing real-time synthetic voice detection with caller identification by voice fingerprint, call recording, and transcription, available to its customers now and reaching resellers in Q4.

Pindrop builds Know Your Agent for the phone line

The most consequential launch may be Pindrop's BotStopper, a standalone product that identifies AI agents and automated callers in roughly two seconds, powered by what the company calls the AI Voice Consortium, a registry of more than 5,000 AI voices in a Pindrop-managed database. CEO Vijay Balasubramanian's positioning is deliberate: the goal is not blocking machine callers but knowing which agent is calling, whom it represents, and what it may do, a phone-channel version of the Know Your Agent problem payment networks are solving for checkout. The claimed stakes are concrete: at one healthcare organization, bot calls peaked at 6,668 in a month with about 15.1 million dollars in HSA balances exposed, and fell 94.3 percent within four months of real-time detection, by Pindrop's own account. The company is also extending continuous video analysis to screen-sharing sessions through Pulse for Meetings.

The pattern: continuous, multi-signal, and humble about humans

Three design commitments run through every launch. Detection is continuous rather than point-in-time, because a caller verified at minute one can hand off to a clone at minute six, the gap that made the Fideuram messaging fraud and the wave of executive impersonation attacks so expensive. Detection is multi-signal, stacking hardware forensics, audio artifacts, and behavioral tells, because single-metric accuracy claims decay fast against new generators. And every vendor has stopped asking employees to be the detector, which is the quiet concession that training people to spot deepfakes was always a losing control.

The buyer's caveats are the usual ones: the detection rates and the healthcare case study are vendor-reported, registry approaches like the Voice Consortium only cover agents that enroll, and a conversation-layer product that is not wired into identity verification can tell you a voice is synthetic without telling you who is actually present. The deepeye-backed approach is to treat conversation-layer signals as one evidence stream among several, fused with document, device, and liveness checks, so the question answered is not just is this voice real, but is this the verified person authorized to do what they are asking.

Live Conversation Deepfake Defense FAQ

What is driving deepfake detection into live conversations?: AI-generated audio and video paired with social engineering now target calls, meetings, and contact centers directly, and vendors have concluded human participants cannot reliably detect manipulation aimed at their psychology, so detection is moving into the channel itself.
What did Netarx launch?: An AI Defense API built on an orchestration layer of 40 specialized models cross-validating signals in real time, spanning device fingerprints, EXIF data, compression signatures, audio micro-artifacts, and video tells such as blink frequency and lighting inconsistencies.
What is Pindrop BotStopper?: A standalone detector that identifies AI agents and automated callers in about two seconds, backed by a registry of more than 5,000 AI voices. Pindrop positions it as Know Your Agent for voice: identify the agent and its principal rather than block all machine callers.
Are the vendor accuracy claims verified?: No. Detection rates, the two-second identification figure, and the healthcare case study with its 94.3 percent reduction are company statements, and registry-based approaches only recognize enrolled voices, so independent testing remains essential.
How should firms combine conversation-layer detection with identity verification?: Use synthetic-media detection as one evidence stream and fuse it with identity, device, and liveness signals, so high-risk actions requested mid-call route to verified, human-bound approval ceremonies rather than resting on a single detector's verdict.