Deepfakes Hit the Contact Center: The New Vishing Threat

Deepfake fraud attempts in contact centers rose more than 1,300% in 2024, from roughly one a month to seven a day, according to the Pindrop 2025 Voice Intelligence and Security Report. The phone line, long treated as a low-risk channel, has become one of the busiest fronts in fraud.

Here is the short answer to why this is happening now: contact centers authenticate callers by voice and knowledge questions, and synthetic voice defeats both. A cloned voice answers the security questions in the account holder’s own tone, and the agent, trained to be helpful, resets the password or approves the change. Vishing, voice phishing, has found its scale.

Why the contact center is the soft target

The contact center is the soft target because it is built to say yes. Agents are measured on call handling time and customer satisfaction, so friction is discouraged by design. Authentication often rests on knowledge-based questions like a date of birth or the last four digits of an account, all of which are cheap to buy on breach markets, plus the sound of the caller’s voice, which can now be cloned.

The financial stakes are large. Pindrop projected $44.5 billion in contact-center fraud exposure for 2025, a figure that reflects how many high-value actions, including password resets, wire approvals and account takeovers, run through the phone channel. Attackers follow the money, and the money flows through the call floor.

There is also an incentive mismatch baked into most operations. Agents who add friction to a legitimate call get worse satisfaction scores, while agents who let a fraudster through rarely face a personal consequence, because the loss surfaces days later in a different department. Until the organization rewards caution as much as speed, the human layer stays tilted toward the attacker.

How a deepfake vishing attack works

A deepfake vishing attack combines stolen data with synthetic voice to pass authentication. The fraudster gathers a target’s personal details from a breach, then clones the target’s voice from a short public clip or a recorded earlier call. When they dial the contact center, they can both answer the knowledge questions and match the expected voice, clearing two checks at once.

The IVR is often the first probe. Attackers run the interactive voice response system to test which stolen credentials work and which accounts are vulnerable, then escalate to a live agent for the high-value request. The move from one attempt a month to seven a day, the 1,300% jump Pindrop recorded across 2024, reflects automation: crews now run these probes at volume rather than one careful call at a time.

The reconnaissance can be patient. A fraudster may place several low-value calls to map an account’s balance, recent transactions and the questions an agent asks, all without triggering a fraud flag, then use that intelligence on a single decisive call. By the time the cloned voice reaches a human agent, the attacker often knows more about the account’s recent activity than the agent does.

Why voice biometrics alone no longer hold

Voice biometrics alone no longer hold because the thing they measure can be synthesized. Legacy voiceprint systems compare a caller’s audio to an enrolled sample of the real customer. A high-quality clone is engineered to match that sample, so the very feature meant to prove identity becomes the feature the attacker reproduces. Static voiceprints were designed for a world where voices could not be faked at scale. That world is gone.

Liveness and deepfake detection close the gap that biometrics leave open. Rather than only asking “does this match the enrolled voice,” modern deepfake detection software asks “is this audio generated by a machine,” analyzing spectral artifacts and synthesis signatures a human ear cannot catch. Pairing that signal with the existing voiceprint restores the layered defense that a clone alone can otherwise defeat.

Building a layered contact-center defense

A layered defense assumes any single check can fail and stacks several so an attacker has to beat all of them. Start with real-time synthetic-voice detection scoring calls in the first seconds, so a machine-generated voice is flagged before an agent acts on it. Add device and network signals, since a spoofed caller ID or a suspicious VoIP origin is a strong risk indicator independent of the audio.

Then harden the process around the agent. Set step-up verification for high-risk actions like password resets, payee changes and large transfers, requiring an out-of-band confirmation the caller cannot fake by voice. Give agents a clear, low-friction path to pause a suspicious call without penalty to their handle-time metrics, because a policy that punishes caution guarantees agents ignore red flags. With $44.5 billion in projected exposure for 2025 and attempts up 1,300%, the cost of one missed call now dwarfs the cost of a few extra seconds of verification.

Frequently asked questions

What is vishing and how is it different with deepfakes?

Vishing is voice phishing, fraud carried out over the phone. Traditional vishing relied on the attacker’s own voice and social pressure. Deepfake vishing adds a cloned voice matching a specific person, so the attacker can defeat voice-based authentication and impersonate a real account holder or executive. That makes the attack far harder for an agent to detect by ear.

Can voice biometrics stop deepfake callers?

Not on their own. Voiceprint systems compare a caller to an enrolled sample, and a good clone is built to match that sample. Attackers reproduce the exact feature the biometric checks. Effective defense pairs voiceprints with synthetic-voice detection, which looks for signs that audio was machine-generated rather than only whether it matches the enrolled voice.

Why did contact-center deepfake attempts rise so sharply in 2024?

Automation and cheaper cloning tools. Pindrop recorded a jump of more than 1,300%, from about one attempt a month to seven a day, as crews began running synthetic voices at scale against IVR systems and agents. Lower tooling costs plus large stores of breached personal data made high-volume vishing practical rather than a one-off effort.

How do agents fit into the defense if they cannot hear a deepfake?

Agents are the backstop, not the detector. Since staff cannot reliably hear a clone, the system flags suspect audio and enforces step-up verification for risky actions. Agents need clear authority to pause a call and escalate without hurting their performance metrics. Removing the pressure to rush is as important as any detection tool.