Your finance manager receives a voice message from the managing director: a supplier problem, a confidential acquisition, a transfer needed before the bank closes. The vocabulary is right. The voice is right. The director really is travelling.
The wrong question is: “Can the finance manager spot that the voice is cloned?”
The right question is: “Can any voice message authorize this payment?”
Deepfake and voice-cloning attacks strengthen an existing fraud pattern: impersonate a person with authority, create urgency, and move the target outside the normal process. The durable defence is procedural. It should work whether the request arrives by email, phone, video meeting, compromised account, or a real executive making a hurried mistake.
[!SECURITY] Identity and authority are separate checks. Confirm who is requesting the action through an independent channel, then confirm that the action itself satisfies the company’s approval rules.
The threat is broader than a fake voice
An impersonation attempt may combine:
- a spoofed or compromised email account;
- information from social media, calendars, invoices, or a breached mailbox;
- convincing text in the sender’s normal language;
- cloned voice notes or live synthetic audio;
- manipulated video or a static profile image in a meeting;
- a second impersonator posing as a lawyer, supplier, bank, or colleague;
- pressure to bypass controls because the matter is urgent or confidential.
The FBI’s AI-enabled cybercrime warning describes voice and video cloning used to impersonate co-workers and business partners. Its current business email compromise guidance recommends independently verifying payment requests and changes in account details. Sources rechecked 28 July 2026.
Do not build policy around one channel or one generation technique. Build it around consequential actions.
Control 1: An independent callback
For any unexpected request involving money, credentials, sensitive records, payroll, or bank-detail changes:
- end or pause the incoming contact;
- open a separate, previously trusted channel;
- contact the requester using a number or account already held in company records;
- verify the exact action: recipient, amount, account, reason, deadline;
- record who verified it and through which channel.
Do not call the number included in the request. Do not let the requester “transfer you to finance.” Do not treat the same meeting’s chat as an independent channel.
For a supplier bank-detail change, call the supplier contact from the master vendor record, not the invoice or change notice. Then require confirmation from your internal vendor owner.
The callback is not a biometric test. It breaks the attacker’s control of the conversation.
Control 2: Two-person approval for consequential changes
No single person — including the CEO — should be able to request and complete a material payment or sensitive account change outside the system.
Define thresholds and actions that require two people:
- new payee or changed bank details;
- payment outside a purchase order or contract;
- urgent or confidential transfer;
- payroll destination change;
- release of customer, employee, or financial records;
- password, MFA, recovery, or administrator change;
- gift cards, cryptocurrency, or unusual payment rails;
- remote-access installation.
The second approver should see the underlying evidence and perform their own check. Forwarding “the CEO confirmed” is not independent approval.
Your accounting or banking platform should enforce the separation where possible. A policy document is weaker when the system still lets one credential complete the action.
Control 3: No urgency override
Create an exception process, not an exception-shaped hole.
The rule can be:
Urgency may shorten the response time. It never removes independent verification, approval thresholds, or bank-detail checks.
If the normal approver is unavailable:
- use a named alternate;
- reduce the allowed amount;
- delay the action;
- or escalate to a documented owner.
Never improvise a new approval chain inside the suspicious conversation.
Confidential matters still need controls. Give finance staff permission to say: “I do not need the background, but I do need the second approval required for this action.”
Put the controls in a table
| Event | Required verification | Approval | Record |
|---|---|---|---|
| Existing supplier, normal invoice | Match purchase evidence | Normal threshold | Invoice and approver |
| New or changed bank account | Callback to master contact | Two people | Contact, time, change, approvals |
| Executive urgent payment | Independent callback | Two people; no urgency bypass | Request, verification, approvals |
| Payroll destination change | Employee callback via HR record | HR plus payroll owner | Change ticket and confirmation |
| Sensitive data release | Requester and lawful purpose verified | Data owner | Scope, recipient, basis, delivery |
| Credential or MFA reset | Approved identity-recovery route | System owner where privileged | Reset and session revocation |
Adapt the thresholds, but do not leave the events undefined.
A five-minute staff briefing
Use a scenario, not a technology lecture:
“A message, call, or video may look and sound exactly like a director, colleague, bank, or supplier. That does not change our process.
“If a request involves money, credentials, sensitive data, or account changes, pause it. Start a new contact using details already in our records. Verify the exact action, then use the required second approval.
“Urgency and confidentiality do not remove those checks. Nobody here will criticize you for delaying an action to verify it. Report suspicious contact to [role/channel] and preserve the message.”
Then ask each participant:
- Where is the trusted supplier number stored?
- Who is the alternate approver today?
- How do you report a suspicious request?
- Can the CEO bypass the control?
If the answers differ, the briefing has found a process defect.
Rehearse the attack
Run a tabletop exercise without creating a deepfake:
- Give finance a fictional urgent request from a travelling director.
- Include a changed supplier account and a confidentiality demand.
- Make the normal approver unavailable.
- Observe which channel they use, what evidence they request, and whether the alternate path is clear.
- Test bank and incident contacts.
- Record gaps and fix the procedure.
There is no need to generate a cloned voice. The purpose is to test the decision process, not the realism of the media.
Repeat after staff or supplier changes and at least annually for workflows that can move material funds.
If an attempt happens
If no money or access moved:
- preserve the original email, message, audio, meeting details, numbers, and headers;
- notify the security or IT owner;
- check whether an account was compromised rather than merely spoofed;
- warn likely targets through a trusted channel;
- review recent payment and account-change activity;
- report the attempt through the appropriate national channel.
If a payment moved:
- Contact the bank immediately and ask it to stop or recall the transfer.
- Contact the receiving institution if your bank instructs you to.
- Preserve evidence and write a timeline.
- Lock or reset compromised accounts, revoke sessions, and review mailbox rules and MFA.
- Report to law enforcement. In Estonia, RIA’s internet-safety guidance directs cybercrime victims to cyber.politsei.ee.
- Notify customers, employees, insurers, regulators, or contractual partners where your incident obligations require it.
Do not let embarrassment delay the bank call. Speed can affect recovery.
What technical controls can and cannot do
Useful supporting controls include:
- phishing-resistant MFA for important accounts;
- restricted mailbox forwarding rules;
- alerts for payment and beneficiary changes;
- role separation in banking and accounting systems;
- domain protection and email authentication;
- monitoring for executive impersonation;
- Content Credentials on official media where supported.
These reduce opportunity and improve evidence. They do not make an incoming voice trustworthy.
Deepfake detectors may help an investigator prioritize evidence, but they should not be the approval gate. A false negative can authorize fraud; a false positive can block a legitimate executive. Keep the procedure independent of detection quality.
Content Credentials and provenance can help demonstrate the origin of your own official media. They do not replace payment verification. The household counterpart is recognising AI-enabled scams.
The honest limit
No control eliminates fraud. A real executive can pressure staff to bypass policy. A compromised supplier can confirm a fraudulent change. Two people can make the same mistake.
The goal is to remove single-message authority, create independent friction, and make stopping normal. That defeats far more than voice cloning. It also reduces ordinary business email compromise, account-change fraud, and expensive errors made under pressure.
Do not train employees to decide whether the voice is real. Give them a process that does not care.
