Deepfake-Proof Payments: Callback Verification That Holds Up

The problem: the request looks right and sounds right

Business email compromise is the older version of this fraud. A criminal takes over, or convincingly imitates, an executive's or vendor's email account and asks for a payment, a changed bank account or a gift-card purchase. Companies learned to look for odd addresses, strange phrasing and bad grammar. Generative AI removes most of those tells.

The FBI's Internet Crime Complaint Center (IC3) published a public service announcement on December 3, 2024 (alert number I-120324‑PSA) titled "Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud." It says criminals can generate short audio clips that imitate a person's voice, and use cloned voices to impersonate individuals. It also says they can generate videos for real-time video calls with fake company executives, law enforcement or other authority figures.

That last point matters for any finance team. The old fallback was "I'll just hop on a quick video call to be sure." If a live video call can be faked, a video call is no longer a verification step. It is another message that needs verifying.

What the regulators have said

FinCEN, the U.S. Treasury bureau that supervises anti-money-laundering reporting, issued alert FIN-2024‑Alert004 on November 13, 2024. Its title is about fraud schemes involving deepfake media targeting financial institutions. FinCEN says it has seen more suspicious activity reports describing suspected deepfake use, particularly fraudulent identity documents used to get around identity verification and authentication methods. The alert is written for banks and similar institutions, but the lesson applies to any company: media that looks authentic is now cheap to produce.

The New York Department of Financial Services (DFS) sent an industry letter on October 16, 2024, titled "Cybersecurity Risks Arising from Artificial Intelligence and Strategies to Combat Related Risks." On the DFS site, the letter says AI-enabled social engineering is one of the most significant threats to the financial services sector, and that attackers have used deepfakes to persuade employees to make unauthorized wire transfers. It suggests that regulated entities consider avoiding authentication by SMS text, voice or video, because deepfakes can imitate those channels, and consider forms of authentication that deepfakes cannot impersonate, such as digital certificates and physical security keys. It also points to simulated voice and video impersonation exercises as a way to train staff. DFS letters apply to DFS-regulated entities; we cite it here as a description of the threat and of one regulator's suggestions, not as a rule for your business.

Taken together, the FBI, FinCEN and DFS describe the same shift. The signal that used to build trust, a familiar voice or a face, can be manufactured. Controls have to move to something the attacker cannot manufacture.

The principle: verify on a channel the requester does not control

A criminal who has taken over a mailbox, cloned a voice or joined a call with a synthetic face controls the channel the request arrived on. Anything you do inside that channel can be intercepted or imitated. Replying to the email, asking the "CEO" on the same video call to confirm, or calling the number in the email signature all leave the attacker in control.

Out-of-band verification means switching to a different channel that you chose, using contact details the requester did not supply. The IC3 announcement gives the consumer version of this advice: verify a caller by hanging up and contacting the organization directly using a number you look up yourself. A business version is the same idea, made routine.

A callback procedure that holds up

A usable procedure is short enough that staff follow it at 4:45 on a Friday. This is the shape we help clients design and document.

• Define the trigger. Any request to send a payment above a threshold you set, any new payee, and any change to bank details, remittance instructions or contact information triggers verification. So does any request marked urgent or confidential, whoever it comes from.
• Call back a number already on file. Use the phone number in your vendor master record, your contract, your last invoice you verified, or your internal directory. Never use a number, link or meeting invitation that came with the request.
• Speak to a person you can identify independently. For a vendor, that means a contact on file, not whoever picks up a number given in the email. For an internal request, call the executive on their directory number, or walk to their office.
• Confirm the details, not just the intent. Read back the amount, the destination account (last four digits are enough to compare) and the invoice or purpose. Ask the person to state them first where practical, rather than confirming what you read out.
• Require a second approver. A different person, with no reporting line to the requester, approves the release after reviewing the verification record.
• Record it. Note who requested, who called back, which number was used and where it came from, the date and time, who approved, and the outcome.
• Hold changes to bank details for a cooling-off period where the business can tolerate it, and send a notice to the old contact details on file.
• Provide an escalation path. If the requester pushes back on verification, staff should be able to escalate to a named person without penalty. Pressure to skip the step is itself a warning sign.

Why the second approver often fails

Dual control is the most common fraud control and the most commonly hollowed out. Two failure patterns appear again and again.

First, the second approver reviews the same forwarded email as the first. If the email is fake, both approvals rest on the same fake. The second person needs to see the verification record, meaning the callback notes, not just the request.

Second, the second approver is junior to the requester or afraid of delaying an executive. If a controller will not question a payment that appears to come from the CEO, the control exists on paper only. Leadership has to say in writing, and mean it, that verification applies to the CEO's requests too, and that staff who follow the procedure will be backed when it slows a payment.

A related point: the approver's identity matters in systems as well as in conversations. Where the payment platform or bank portal supports it, require approvals from separate authenticated users with phishing-resistant sign-in, such as passkeys or hardware security keys, so a stolen password alone cannot approve a payment. Our article on multi-factor authentication and passkeys covers that layer.

Where callback procedures break in practice

• The number on file was changed by the attacker. If a criminal who controls a mailbox also updated the vendor record, the "number on file" is no longer trustworthy. Limit who can edit vendor master data, log every change, and verify contact-detail changes through a previously known channel.
• SIM swaps and call forwarding. A phone number is not an identity. For very high-value transfers, add a second verification factor, such as an in-person confirmation or a pre-agreed internal process on a separate platform.
• Callbacks done by the person who received the request. If the employee who got the email also makes the call, tired habit takes over. Where possible, the caller should be someone who did not receive the original message.
• Verbal shared secrets that leak. The IC3 announcement suggests agreeing on a secret word or phrase with family members. A similar code phrase can work inside a small company, but treat it as a supplement. It does not replace calling back, and it must be changed when staff leave or it may have been exposed.
• Exceptions that become the rule. If the procedure is skipped for "good customers" or end-of-quarter rushes, criminals will aim at those moments. Log each exception and review the log monthly.

Do not rely on spotting the fake

The IC3 announcement lists visual flaws to look for, such as distorted hands, irregular faces, unnatural shadows or movements, and lag, and says to listen closely to tone and word choice. These are fair tips for staff awareness. They are not a control. Detection tools and human judgment are both imperfect, and quality improves faster than training materials age. We do not tell clients that any tool or training will catch every deepfake, and neither should a vendor. The procedure is designed so that it still works when the fake is perfect.

Test it: drills and tabletop exercises

A procedure is only as good as the last time someone used it. DFS's October 2024 letter points to simulated voice and video impersonation exercises as a training method. A drill does not need a synthetic face to be useful.

• Run a scheduled test. With leadership approval, have a trusted person send a realistic payment-change request to the finance team, or place a call impersonating a vendor. Check whether staff follow the callback steps and whether the second approver asks for the record.
• Run a tabletop exercise twice a year. Walk through a scenario: the CFO is travelling, an urgent wire is requested by video call, the bank cut-off is in an hour. Who can approve? Who calls whom? What if the callback goes to voicemail?
• Measure the result as a process, not as blame. Record how long verification took, where staff improvised, and what to fix. Reward the person who stopped the test payment.
• Update the procedure after each drill, and keep the revised copy where finance staff can find it quickly, including on a day when email is down.

Insurance: check your policy with your broker

Some crime and social-engineering insurance policies condition cover on the insured following a verification procedure for payment instructions, and some limit or exclude losses when staff did not follow it. Wording differs between policies. We cannot tell you what yours says or whether a particular loss would be paid. Check your policy with your broker, ask what verification steps are required, and ask what evidence they expect you to keep. A written, tested procedure and a verification log are the kind of evidence worth having on file either way. This article is not insurance advice.

What Clayton Cybernetics does here

Our team helps clients write the payment verification procedure in plain language, map it onto their accounting and banking workflows, and set up the technical layer behind it: phishing-resistant sign-in for finance accounts, mailbox rules and logging that make account takeover visible, and restricted, logged access to vendor master data. We design and run simulation exercises with staff. We assemble the written procedure, the drill results and the verification log into a file your broker or auditor can review; they draw the conclusions. If you want an outside view of how a fake payment request would travel through your company today, start with a free security assessment.

Sources

• FBI Internet Crime Complaint Center: Public Service Announcement I-120324-PSA, Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (December 3, 2024)
• FinCEN: FIN-2024-Alert004, Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions (November 13, 2024), news release
• FinCEN: FIN-2024-Alert004, full alert (PDF)
• New York DFS: Industry Letter, Cybersecurity Risks Arising from Artificial Intelligence and Strategies to Combat Related Risks (October 16, 2024)

This article is general information, not legal or insurance advice. It reflects the sources listed as of November 8, 2026. Insurance policy terms differ; check your policy with your broker. Confirm your obligations with your own counsel. Last reviewed November 8, 2026.

Keep reading

Next · RegulationWhat Insurers and Auditors Now Ask About Your AIQuestions about AI are arriving in insurance renewals, vendor questionnaires and examinations. They tend to cluster around six things: a list of the AI you use, what each agent is allowed to do, what gets logged, where a human must approve, what testing has been done, and who is responsible. This article shows where those six come from. It draws on the Cloud Security Alliance's AI Controls Matrix, ISO/IEC 42001, the NIST AI Risk Management Framework and FINRA's 2026 Regulatory Oversight Report, and it describes what each one actually says. It also covers a change in general liability insurance forms that many businesses have not noticed. It is a map of the questions, not a compliance verdict.9 min read

• Penetration Testing Explained for Executives: Scope, Reports and How Often to Test8 min
• Ransomware Readiness and Incident Response: A Checklist Based on the CISA #StopRansomware Guide9 min