Deepfakes and the payment-approval call
By Sam Rivera, Founder, SentinelPanda · October 24, 2026 · 4 min read · AI Governance
The control was never the approval workflow. It was that you knew what your CFO sounded like.
The control nobody wrote down
Ask a finance team how an urgent payment request gets verified and you will usually hear about an approval threshold, a second signatory, and a form. Watch what actually happens and the decisive step is earlier: somebody recognised the voice on the call, or the face in the meeting, and the rest of the process proceeded on that basis.
That recognition is a real control. It is just not in any document, which means nobody noticed when it stopped working.
What changed
Synthetic voice now needs very little reference audio, and a recorded earnings call, a conference talk or a podcast appearance supplies more than enough. Live video is harder but no longer exotic. The result is that an attacker can present as a specific, known senior person rather than as a plausible stranger.
This matters more for payment fraud than for data theft, because the payment path is short. A convincing call to someone with authority to move money can complete in a single conversation, with no malware, no credential theft and no lateral movement for anyone to detect.
Why awareness training misses it
Requirement 12.6.3 asks for security awareness training at least annually and upon significant change, and 12.6.3.1 asks that it include threats and vulnerabilities that could affect the security of the cardholder data environment — phishing and related attacks, and social engineering, named explicitly.
Most programmes satisfy this with email content. Staff are taught to check sender addresses, hover over links and distrust urgency in writing. None of that transfers to a phone call from someone they recognise. The training is defensible against the clause and useless against the attack, which is the worst combination because it reads as covered.
What actually replaces recognition
- Call back on a number from your own directory. Not a number offered during the call, not one in the signature of the email that prompted it.
- Make it unconditional. A verification step that can be waived for seniority or urgency will be waived precisely when it matters, because those are the pressures the attack manufactures.
- Give juniors explicit cover. The attack works on hierarchy. Staff need to know in advance that pausing a payment to verify is correct behaviour and will be supported afterwards, whoever was on the call.
- Verify the change, not the request. Fraud usually arrives as an amended bank detail or a new beneficiary rather than an invented invoice. Treat a change to payment details as the event requiring verification.
- Drop shared secrets. A passphrase that two people agree on is fine until it is in an email, and it usually is.
The card-specific version
The same pressure appears against customer-facing staff. A caller presenting as a customer, or as an internal manager, asks an agent to read back a card number or to confirm a security code. The social engineering is the mechanism; the compliance failure is what the staff member does next.
Requirement 3.3.1 prohibits retaining the security code after authorisation, so the correct answer is that it cannot be read back because it is not held. That is a much easier instruction to follow under pressure than a judgement call about whether the caller sounds legitimate, which is the point — procedural answers survive stress better than perceptual ones.
Detection is not the strategy
There is real work in synthetic media detection, and it will improve. It is a poor foundation for a payment control, because it makes the defence a perceptual contest that the attacker gets to iterate on and your staff do not.
The durable answer is the same one that worked before any of this existed. Verify through a second channel that the attacker does not control, and make the step mandatory so it does not depend on anyone feeling suspicious on the day.