Voice deepfakes are turning contact centers into fraud channels

At a glance

What this is: NHI Mgmt Group's analysis, drawing on Daon's blog post: Voice deepfakes are pushing contact centers into a fraud model where knowledge-based checks and human judgement are no longer dependable controls.

Why it matters: This matters because contact centers sit at the junction of human identity verification, account recovery, and high-risk approvals, so weak caller authentication can become a direct fraud path into customer accounts.

Context

Contact center fraud is no longer just a customer-service problem. It is an identity verification problem, because the channel is built to resolve urgent requests quickly while fraudsters exploit that same speed and trust to impersonate legitimate callers.

Traditional knowledge-based authentication breaks down when the answers are already circulating in breach data, data broker records, or dark web markets. Deepfake audio adds a second failure mode by making synthetic callers sound plausible enough that both agents and scripted checks can be bypassed, which makes this a structural governance issue rather than a training issue.

Key questions

Q: What should teams do first when voice deepfakes are used in contact center fraud?

A: Treat the event as an identity proofing failure and move high-risk callers to stronger controls such as callback to a known number, knowledge-based authentication, or out-of-band verification. Preserve the audio, capture timestamps, and escalate the case so fraud operations can correlate the call with account changes, reset requests, or payment instructions.

Q: Why do voice deepfakes bypass human call center agents so easily?

A: Voice deepfakes succeed because agents tend to trust familiar speech cues, urgency and context more than a caller’s true identity. They exploit human perception by mimicking tone, accent and conversation flow, so the call sounds legitimate even when the speaker is synthetic. Stronger controls must verify the caller through out-of-band checks, not voice alone.

Q: How do passive voice biometrics work in contact centers?

A: Passive voice biometrics analyse a caller’s live speech and compare it with an enrolled voiceprint to estimate whether the voice matches a known identity. Contact centres use it as one signal within risk-based authentication, often combined with device, account, or knowledge checks. It can reduce repeated challenge questions, but enrolment quality and replay attacks matter.

Q: How should organizations verify high-risk caller requests without adding friction?

A: Use risk-based call handling, so routine requests pass with minimal checks while high-risk changes trigger step-up verification such as a one-time passcode, call-back to a registered number, or agent supervisor approval. The control should be tied to caller risk, request type, and account history, not every interaction, to limit friction and fraud.

Technical breakdown

Why knowledge-based authentication fails in the contact center

Knowledge-based authentication depends on secrets that callers can remember and agents can verify quickly, but those secrets are often discoverable through prior breaches or third-party data sources. Once an attacker has enough personal data, the challenge-response model becomes a weak proxy for identity. In a high-volume contact center, that weakness compounds because the control is designed for speed, not resistance to impersonation. The result is a channel where the authentication factor is often easier to harvest than the account itself.

Practical implication: Treat KBA as a legacy friction control, not a trusted identity proofing method, for any workflow that can move money or reset access.

How passive voice biometrics and liveness detection work together

Passive voice biometrics compares the acoustic characteristics of a live caller against an enrolled voiceprint while the conversation is already under way. Liveness detection adds a separate check for replay or pre-recorded audio, and real-time synthetic speech detection looks for machine-generated voice artifacts in the stream itself. The combination matters because deepfakes can mimic tone and cadence well enough to fool a listener, but not necessarily the underlying biometric and signal-integrity checks. Continuous authentication extends that verification beyond the first few seconds of the call. Read Voice biometrics and voiceprint authentication in the age of voice cloning.

Practical implication: Use layered voice controls so the session stays authenticated after initial verification, not just at greeting.

Why step-up verification is the right control for high-risk requests

Step-up verification changes the trust point from the call itself to a second channel that can confirm the same person when risk rises. That matters when a caller asks for a password reset, payee change, or high-value transfer, because those are the moments where impersonation becomes financially material. The architecture only works when the contact center has a single customer identity record across channels, so that voice anomalies and device anomalies are evaluated against the same profile. Without that unification, the control becomes inconsistent and easier to evade. See the Human-in-the-Loop for AI Agents Guide.

Practical implication: Gate sensitive transactions behind a second factor that is bound to the enrolled customer identity, not the call alone.

Threat narrative

Attacker objective: The attacker wants to impersonate a customer well enough to obtain account control, authorise sensitive actions, or trigger financial transfers.

1. Entry begins with synthetic audio or a cloned voice entering the call stream and presenting as the customer.
2. Credential access occurs when the fraudster uses social engineering and weak knowledge-based checks to clear initial authentication.
3. Escalation follows when the attacker requests account changes, reset actions, or payment instructions that convert identity deception into account control.
4. Impact is fraud loss, unauthorised account activity, or downstream customer compromise after the caller is accepted as legitimate.

Breaches seen in the wild

• Arup deepfake fraud 2024: Deepfakes of Arup's CFO and colleagues on a video call led a Hong Kong employee to transfer HK$200 million (about US$25.6m) to fraudsters.

Read and download The State of NHI & AI Agent Breach Report 2026, covering 200+ breaches impacting Non-Human Identities including AI Agents.

NHI Mgmt Group analysis

Voice deepfakes expose an authentication design flaw, not just a fraud trend. Contact centers were optimised for speed and courtesy long before synthetic speech became accessible at scale. That makes the channel structurally vulnerable wherever the control model depends on memory, caller confidence, or agent intuition. The governance lesson is that identity proofing must move from questions to biometric and session-level evidence, because the old model was never built for machine-generated impersonation.

The contact center now sits inside the human identity attack surface. Fraudsters are not attacking the phone channel in isolation. They are chaining voice cloning, breached personal data, and high-friction customer recovery flows into one abuse path that turns service interactions into account takeover opportunities. Deepfake impersonation controls become relevant here because out-of-band verification and transaction gating are now part of the identity model, not optional extras.

Continuous authentication is becoming the baseline for trustworthy voice channels. A one-time match at the start of a call is no longer enough when an attacker can change speakers, switch audio sources, or hand off the call after initial clearance. The real control gap is session continuity, not first-contact verification. That shifts the governance question from whether a caller sounded right at the door to whether the full interaction remained bound to the same verified identity.

Deepfake resistance should be treated as a customer trust control, not a niche antifraud feature. The organisations most exposed are the ones that still separate fraud prevention, IAM, and contact center operations. Voice-based abuse crosses all three domains, so the control strategy has to connect identity data, device history, and transaction risk in one decision path. Practitioners should treat this as a shared governance problem, not a channel-specific tweak.

From our research library:

• Veriff's 2026 Identity Fraud Report documented a 300X increase in digitally presented media that was either entirely AI-generated or otherwise altered.

What this signals

Identity proofing is shifting from static facts to live behavioural evidence. Contact centers that still anchor trust in security questions will keep losing ground to synthetic audio and breached personal data. The more durable model is one that treats the caller’s voice, device, and session behaviour as a single verification problem, because fraud now crosses those signals in one flow.

Deepfake-ready fraud operations are becoming cross-functional by necessity. Fraud teams, IAM teams, and contact center owners can no longer operate separate trust models for the same customer journey. The next governance gap is coordination, because attackers exploit every seam between channel controls, identity records, and transaction approval rules.

For practitioners

• Replace KBA on high-risk calls Remove security questions from any workflow that can reset credentials, change payees, or release funds, and substitute identity checks that are harder to precompute from breached data.
• Add liveness and synthetic speech checks Deploy controls that can detect replay audio, cloned voices, and machine-generated speech during the live call, not only at enrolment or post-call review.
• Extend verification beyond call start Require continuous authentication or step-up checks on every high-risk branch of the call flow so an attacker cannot clear the front door and change identity mid-session.
• Unify voice and device identity signals Use a single customer record that joins telephony events, enrolled voiceprints, and registered-device history so anomalies can be evaluated in the same decision engine.
• Escalate suspicious calls to out-of-band confirmation Route materially risky requests to a mobile app or other secondary channel before any irreversible action is approved, especially when caller behaviour or audio quality changes unexpectedly.

Key takeaways

• Voice cloning has turned contact centers into a practical impersonation channel, especially where recovery and payment workflows still depend on caller knowledge.
• The weakest point is not only the audio itself but the governance model that trusts a call before it is tied to a verified identity.
• Controls that combine biometric matching, liveness detection, continuous verification, and out-of-band step-up checks reduce the attack window materially.

Related NHI Mgmt Group resources: Deepfakes, Social Engineering and AI Impersonation Guide · Voice biometrics and voiceprint authentication in the age of voice cloning · Human vs Non-Human Identity · Human-in-the-Loop for AI Agents Guide

Key terms

• Voice Deepfake: A voice deepfake is synthetic or manipulated audio designed to imitate a real person’s voice. Attackers use it to impersonate a user during authentication or to support social engineering. Defending against it requires liveness testing, anomaly detection, and risk-based authentication decisions that do not trust audio alone.
• Knowledge-based authentication: A verification method that asks a person to prove identity by answering facts such as past addresses, account details, or other shared information. It is operationally convenient but weak against social engineering, breach data, and impersonation, which is why it performs poorly in high-risk support flows.
• Passive Biometrics: Passive biometrics uses background behavioural traits, such as typing rhythm or mouse movement, to support ongoing identity confidence without forcing a separate challenge. It is useful in fraud detection because it adds continuous context, but it still needs calibration to avoid false alarms.
• Continuous authentication: A model where access is re-evaluated after the initial login instead of being trusted for the full session. It uses live signals such as posture, telemetry, and policy to detect when a session should be stepped up, constrained, or revoked.

Deepen your knowledge

NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, secrets management, and workload identity alongside the identity controls that matter in modern programmes. It gives identity and security practitioners a common language for governing service accounts, credentials, and access paths across the enterprise.