Blog/AI & the future

Voice Cloning and the End of Recognition as a Verification Method

Recognizing a voice or face was never authentication, and cloning has made that obvious. How to find the recognition-dependent controls you still rely on.

CyberWatch AILast reviewed 21 November 2026 · 6 min read
A person holding a mobile phone to their ear while standing beside an office window

Recognition was never authentication. We treated it as if it were, because for most of human history the cost of imitating a specific person's voice well enough to fool their colleague was prohibitively high. That cost has collapsed, and a large number of processes built on the old assumption are still running.

This is a different argument from "watch out for deepfake calls". The point is not that a new attack exists. The point is that an entire class of control, the kind whose description contains "they will know if it is not me", has quietly stopped working, and most organizations have never made a list of which of their controls are in that class.

The inventory nobody has made

Spend twenty minutes writing down every process in your organization where someone's identity is established by a person recognizing them. The list is usually longer and more alarming than expected.

  • The help desk resetting a password or re-enrolling multi-factor authentication for a caller who sounds like the right person and knows their own job title.
  • Payment approval where the second authorizer confirms by phone and the familiar voice is the confirmation.
  • Out-of-hours exceptions, which almost always collapse to recognition because the normal route is closed.
  • Supplier bank detail changes, confirmed by calling a number from the email that requested the change.
  • Voice biometrics on a customer line, where repeating a phrase grants account access.
  • Internal escalation, where a senior person's voice on a call ends a debate about whether something is authorized.
  • Remote onboarding, where a video interview establishes that the person who will receive credentials is the person who applied.
  • Family and personal cases, including the grandparent scam, which is the purest form of recognition as authentication.

Each of these was reasonable when imitation was hard. The FBI's Internet Crime Complaint Center has documented campaigns using AI-generated voice messages to impersonate senior officials and establish a relationship before delivering a malicious link, and NSA, FBI and CISA published joint guidance on deepfake threats to organizations whose recommendations are about verification and preparation rather than about teaching staff to detect artifacts. That emphasis is the tell. The agencies closest to the problem are not advising better listening.

Why "train people to spot it" fails

Three reasons, each sufficient on its own.

The artifacts are a moving target. Any specific tell you teach, a flat affect, odd breathing, a slight delay, will be absent from the next generation of tools. Training built on current artifacts is obsolete before the next refresher.

Conditions destroy the signal. Phone audio is heavily compressed. Add background noise, a poor connection, and a caller who sounds rushed, and the subtle qualities you trained on are gone. Attackers know this and will explain the bad line for you.

Detection is the wrong task to assign. Asking a junior staff member to adjudicate whether the chief executive's voice is real, under time pressure, with a financial consequence either way, places an unreasonable burden on the person least able to carry it. The failure is a design failure, not a training failure.

There is one thing training should do: lower confidence. A member of staff who believes a familiar voice proves identity is more dangerous than one who knows it proves nothing. A no-blame reporting culture matters here, because the behavior you want, pausing an urgent request from a senior person, is socially expensive unless you make it explicitly expected.

What replaces recognition

The replacements share one property: they do not require anyone to judge authenticity.

Channel control

Verification happens through a channel the organization chose, not the one the request arrived on. In practice: hang up, call the number in your own directory. Not the number in the email, not the number the caller gives you, not the number on the caller ID display, which is trivially spoofed. Our walkthrough on verifying a caller is the version to hand to staff.

Process matching

Define the route for each sensitive action, and make arrival outside that route the trigger rather than anything about how the request sounded. A bank detail change that comes in by phone is anomalous because of the channel. Nobody has to assess the voice.

Dual authorization with independent confirmation

Payments above a threshold, and every change to supplier banking details, require a second approver who contacts the supplier on a number from your own records. The written form of this is in our payment verification procedure, and the reason to write it down is that undocumented controls become whatever the busiest person decides at the time.

Cryptographic authentication for systems

Where the thing being protected is system access rather than a payment, the answer is phishing-resistant multi-factor authentication. Passkeys and hardware security keys do not care how convincing the caller was, because there is nothing for the caller to talk someone out of. CISA's guidance on implementing phishing-resistant MFA is the practical reference, and enforcing two-step verification covers the rollout.

Pay particular attention to MFA re-enrollment at the help desk. It is the most common place where a strong cryptographic control is undone by a weak conversational one. Requiring an out-of-band confirmation, a manager's approval, or in-person collection for re-enrollment closes the gap.

Pre-agreed shared phrases, used narrowly

For a small number of specific relationships, a phrase agreed in advance and never written down anywhere searchable is a cheap and surprisingly effective control. It works for a chief executive and a finance director, and it works inside a family for the emergency call scenario. It does not scale to an organization, and it fails if the phrase lives in a shared document.

Got a message or link you are unsure about? Paste it into CyberWatch AI Scan for a free check.

Reducing the raw material, honestly

You will see advice to limit public audio of executives. It is worth a sentence and not more. Executives speak at conferences, record welcome videos and appear on podcasts because that is their job, and a voicemail greeting is enough material anyway. Treat reduction as marginal and spend the effort on the processes instead. The one genuinely useful variant is reviewing whether your own marketing publishes long, clean, single-speaker audio of the people with payment authority, because that is the highest-quality training material an attacker can get for free.

A prediction, labeled as such

Prediction: within three years, voice as a standalone authentication factor will be withdrawn from most consumer financial services, and remaining uses will be as one risk signal among several. The reasoning is that the attack is cheap, the liability sits with the provider, and regulators ask about it.

What would falsify it: major retail banks continuing to offer voice-only account access as a supported primary authentication route in three years, without an additional factor or transaction risk check behind it, and publishing fraud data showing no material increase in losses through that channel. If that happens, the economics are different from what this argument assumes.

A second prediction, held more loosely: real-time video impersonation in interactive calls will become routine in targeted fraud rather than remaining notable. Falsification condition: three years from now, incident reporting from national agencies still treats live video impersonation as an exceptional case rather than a recurring technique, and camera-on remains an effective challenge in practice.

The shift to make internally

The sentence worth repeating to a leadership team is short. We are not going to get better at telling whether a voice is real. We are going to stop needing to know.

That reframing turns an unwinnable detection problem into an ordinary process design problem, and process design is something organizations are good at. The work is finding the controls that still depend on recognition and replacing each one, which is a finite list you can finish. The broader picture of what AI changes and what it does not is in our guide to AI-powered cyberattacks, the organizational playbook for impersonation attempts is in deepfake and voice-cloning fraud against organizations, and the reason not to invest in detection tooling is set out in detecting AI-generated text and images.

Frequently asked questions

How much audio does it take to clone a voice?

Short samples are enough for a recognizable result, and the sample is usually already public: a conference recording, a podcast appearance, a voicemail greeting, a marketing video. Treating the length of available audio as the control has not worked for some time.

Is voice biometrics for bank access still safe?

Voice as the sole factor is a weak design, and some providers have moved away from it. As one signal among several, combined with device, behavior and transaction risk, it still contributes. The problem is any system where saying a phrase is sufficient.

Can you hear the difference in a cloned voice?

Sometimes, and not reliably, and less so each year. More importantly it does not matter: a control that depends on a stressed person on a bad line noticing subtle artifacts is not a control. Build processes that work even when the fake is perfect.

What replaces recognition at a help desk?

Verification against something the organization holds and the caller must produce, plus a callback to the directory number, plus an out-of-band confirmation for high-impact actions like MFA re-enrollment. Never anything the caller can supply from public information.

Does asking a personal question help?

Only if the answer is not discoverable and not guessable, which rules out most of what people ask. A shared phrase agreed in advance with a specific individual works better, because it was never published anywhere.

What about video calls?

Live video is currently harder to fake convincingly than audio, so asking for camera-on raises the attacker's cost. Treat it as friction rather than proof, and keep the callback requirement for anything consequential.

Should we tell staff about voice cloning at all?

Yes, but to lower their confidence rather than to raise their detection skill. The message is that a familiar voice is no longer evidence of identity, and that pausing to verify is expected and carries no social cost.

What single change helps most?

A written rule that no payment, bank detail change, credential reset or access grant is ever completed on the basis of a voice call alone, confirmed through a channel the organization chose. It works regardless of how good the impersonation was.

Sources

  1. Senior US Officials Impersonated in Malicious Messaging Campaign, FBI Internet Crime Complaint Center
  2. NSA, FBI, and CISA Release Cybersecurity Information Sheet on Deepfake Threats, CISA
  3. FTC Announces Winners of Voice Cloning Challenge, Federal Trade Commission
  4. The near-term impact of AI on the cyber threat, UK National Cyber Security Centre
  5. SP 800-63-4, Digital Identity Guidelines, NIST
  6. Implementing Phishing-Resistant MFA, CISA
  7. Internet Crime Report 2024, FBI Internet Crime Complaint Center
  8. 10 Steps to Cyber Security, UK National Cyber Security Centre
  9. ENISA Threat Landscape 2025, ENISA
Share this article
LinkedIn WhatsApp X
For organizations

Your people get these messages at work too.

CyberWatch AI sends your staff realistic practice attacks, trains the gaps it finds, and shows management exactly where the organization stands.