Chapter 03 · Preparing for and responding to deepfake events
When
eyes fail
It now takes about ten seconds of public audio to fake a voice well enough to fool a colleague. You will not out-look the forgery. But you can out-process it, with a callback and a code word.
Four ways reality gets faked
Every deepfake incident is one of four objectives. They need different preparation, different detection, and a different response - and only one of them is really about money.
The forged voice of power
Attackers clone a CFO - the one voice with standing to order a transfer - and direct a mid-level employee to move money under urgency and secrecy. The employee does not lack judgement. They lack a procedure that survives authority.
A finance worker paid out $25 million after a video call with a deepfaked "CFO". A UK firm lost $243,000 to a single cloned voice.
The hire who was never human
A synthetic face and voice pass a remote interview, and a real employee badge follows. This is initial access obtained through HR rather than the network - and the access is fully legitimate once granted.
Campaigns attributed to the DPRK used synthetic personas to land remote developer roles in Western firms, then deployed malware from inside.
The helpdesk that opened the gate
A cloned voice clears human verification. The helpdesk resets the password, enrols the new device, and grants access - to someone who is not there. Voice is not an authentication factor and never was.
A journalist broke into his own bank account using an AI-generated voice, defeating the biometric check that was supposed to replace a password.
The lie that moved markets
A fabricated video of a leader saying the unthinkable - defaming a company, moving a share price, destabilising a public. Google DeepMind ranks mis- and disinformation as the leading malicious use of generative AI.
A fake image of an explosion at the Pentagon briefly erased a reported $500 billion from the S&P 500 before it was debunked.
The fake never had to be perfect
Deepfakes are not defeated by sharper eyes. They are defeated by sequence. Scammers rarely win because the forgery was flawless - they win because they trigger urgency, authority, and fear, which is exactly when checking feels rude and slow.
Studies consistently show that trained people perform barely above chance at spotting deepfakes, and that confidence rises faster than accuracy. Detection technology is a cat-and-mouse game the defence is currently losing. So the answer is not "look harder". It is: follow the procedure, no matter how convincing the face is.
Spot the tells
Three situations, six details each. Select everything that should stop you, then check. The point is not to score well - it is to notice how many of the real tells are about behaviour, not pixels.
The red-flag drill
Select all that apply, per scenario1 · A video call from your CFO
She asks you to authorise an urgent payment to a new supplier, today, before the quarter closes. The video is slightly soft.
2 · A panicked voicemail from "your mother"
An unknown local number leaves a voicemail: she is in trouble, needs £2,000 transferred now, and please do not tell your father.
3 · A final-round remote interview
The candidate is strong on paper and fluent on the call. Something about the video feels off but you cannot name it.
Trust the process, not your eyes
Six controls. None of them require technology, and all of them are boring - which is why they work.
Separation of duties and dual authorisation
No single person both authorises and executes a payment. Two independent approvers for anything significant, with justification chains that cannot see each other's reasoning.
The code of the day
A rotating phrase, delivered through a separate channel with multi-factor protection, that must be spoken to authorise anything. It changes several times a day and can be revoked instantly.
Call back on a known number
Verify any unusual request on a different, pre-registered channel - the number in your contacts, not the one in the message. This one habit defeats almost every voice-clone scam.
Empower the challenge
Give people explicit permission to question a senior leader's request, and make "let's move this to another app" a red flag in itself, not a convenience.
Interview verification
Camera on, no filters or blurred backgrounds, screen shared, at least one in-person touchpoint - and probationary access for every new hire, however good the interview was.
Plan for the event
Write down who monitors, who owns takedown, who speaks publicly, and which forensic retainer you will use. Then rehearse it once, before you need it.
Two scripts worth stealing
The hardest part of verification is the awkward ten seconds where you imply a colleague might be a criminal. A script removes the awkwardness - you are following the procedure, not accusing anyone.
The callback script
"Thanks - I can do that. Our policy is that I verify anything like this on a number I already have, so I'll ring you back on the number in my contacts in the next five minutes. If it's genuine it'll take thirty seconds."
If the request is real, the caller will be relieved. If it is a deepfake, this is the sentence that ends the attack.
The challenge script
"Good - please challenge me on this. If someone asks you to move money after a call from me, call me back on the number you have and ask for the code of the day. I would rather be asked ten times than have this go wrong once."
Say this in a team meeting once a quarter. Permission to challenge decays unless it is renewed out loud.
The code word that actually works
A family or team code word only works if it is unguessable, unused, and never written in the place a scammer would look. Follow three rules and it is the cheapest security control you will ever deploy:
- Choose something that is not on social media, in your email history, or in a shared document.
- Never say it in a voicemail, a group chat, or an email. If it has been typed, replace it.
- Agree a rule for what to do when it is missing: stop, hang up, call back on a known number.
Sources
- OWASP GenAI Security Project - Guide for Preparing and Responding to Deepfake Events - genai.owasp.org
- Figures cited ($25M Arup transfer, $243K UK voice-clone loss, Pentagon-image market dip) come from public reporting; market-move numbers are estimates and vary by source.