Your threat model has a video call in it now: deepfake payment fraud and controls that don't need detection

Disclosure: I co-founded Mitosis Labs, which runs Private SuperIntelligence. This is an adaptation of the original guide, written for family offices.

If you build or secure anything that moves money, here is a definition worth reading closely. The FBI's Internet Crime Complaint Center describes business email compromise as fraud carried out "by compromising email accounts and other forms of communication such as phone numbers and virtual meeting applications."

Virtual meeting applications. The video call is now in scope as an attack channel, not a verification step. The FBI's December 2024 public service announcement is more specific: criminals "generate videos for real time video chats with alleged company executives."

The case that proves the point

CNN reported in May 2024, citing Hong Kong police, that a finance worker in Arup's Hong Kong office received an email that appeared to come from the UK office, asking for a secret transaction. He suspected phishing, which was the right call. Then he joined a video call with people who looked and sounded like the CFO and colleagues he knew. All were deepfakes. He sent HK$200 million (about $25.6 million) across 15 transactions.

Arup told CNN that "fake voices and images were used" and "none of our internal systems were compromised." Rob Greig, Arup's global chief information officer, said the company faces "invoice fraud, phishing scams, WhatsApp voice spoofing, and deepfakes," and that their number and sophistication "has been rising sharply in recent months."

Read that as an incident report. No credential theft, no malware, no lateral movement. The human detection layer worked once (he doubted the email) and was then defeated by a richer signal. Any control that depends on a person judging authenticity has the same failure mode.

Scale, from the source data

From the IC3 2025 annual report:

FinCEN's alert FIN-2024-Alert004 (November 2024) says criminals "have reportedly used GenAI tools to target companies by impersonating an executive or other trusted employee and then instructing victims to transfer large sums."

Family offices sit inside this target group. Deloitte's Family Office Cybersecurity Report 2024 found that 93% of family offices that had been attacked had received phishing emails, and nearly a quarter had faced social engineering.

Controls that are independent of detection

The design principle: authorization must not depend on the authenticity of the requesting channel. Each of these holds against a perfect fake.

• Out-of-band callback to a pre-registered number. The verifier initiates contact on a number already in your records. Never a number, link or invite carried in the request. The FBI gives the same advice.
• Payee changes are new payees. Any change to bank details triggers the callback and a holding period, regardless of sender.
• Two-person release. One prepares, a second approves and confirms the callback occurred.
• Shared secret for humans. The FBI suggests a family code word. Keep it off email and messaging apps.
• Secrecy as a hard stop. A request to hide a payment from the controller is policy-defined grounds to halt.
• Video is not verification. Faces and voices can be generated. Verification happens on a separate channel you started.
• Phishing-resistant MFA. FinCEN lists MFA, including phishing-resistant MFA, as a best practice against deepfake identities.

FinCEN's red flags are written for banks, but two port well: the counterparty asks to switch channels mid-verification, or blames "technical glitches" for a camera or line that keeps failing. Both should terminate the session and trigger a callback.

Five questions before the next wire

1. Do we call back on a number already in our records, never one in the request?
2. Is a change of bank details treated as a new payee with a callback and a holding period?
3. Do two people release every wire, with the approver confirming the callback happened?
4. Have we agreed a family code word in person, and is it kept off email and messaging apps?
5. Is a request for secrecy, or a video call used as verification, written as a stop?

Incident response

The IC3 says "time is of the essence." Contact the bank and request a recall, then file at ic3.gov with full transaction details. In 2025 the IC3's Recovery Asset Team worked 3,900 incidents through its Financial Fraud Kill Chain and froze $679,013,183, a 58% success rate.

Where an AI assistant belongs in this

If you are wiring an assistant into a payments workflow, scope it to read, compare and draft: match a capital call to its commitment and the bank details on file, flag requests that arrive with new instructions, draft the callback note. Do not grant it permission to release payments. Private SuperIntelligence works that way: it acts within permissions the office sets, so a payment can stay a draft until a person approves it.

Speak with the founders

Private SuperIntelligence takes a limited number of families at a time so every one gets the same standard of service. Book a private call (ref DFW1-DEV) to talk through your office with Alex Morris.

Prakshal Jain (PJ), Co-founder and CTO, Private SuperIntelligence (by Mitosis Labs)

Alexander Morris, Co-founder and CEO, Private SuperIntelligence (by Mitosis Labs)

Sources (snapshot 2026-10-24, primary fetches 2026-10-17; recheck on publish day 2026-10-27): 2025 IC3 Annual Report; FBI IC3 PSA, Dec 3, 2024; FinCEN FIN-2024-Alert004; CNN, May 16, 2024; Deloitte Family Office Cybersecurity Report 2024.

Originally published at privatesuperintelligence.si.