Professor Phish
PROFESSOR PHISH — PRE-BRIEF

"Before a face swap operation begins, the attacker doesn't touch a GPU. They do what every intelligence agency does first: harvest the target's face. Watch this unfold in real time."

TARGET IDENTIFIED

Harvesting a Face

The target is Sarah Chen, CFO of NeoVault Financial. The attacker has spent 4 minutes on LinkedIn. That is all it takes.

Sarah Chen
linkedin.com/in/sarahchen-cfo
· 3rd connection · NeoVault Financial

Sarah Chen

Chief Financial Officer at NeoVault Financial

San Francisco, CA · 2,847 connections

TARGET ✓
1
face harvested
2,847
connections
4m
time to harvest
MODEL TRAINING — DeepFaceLive v3.1
Face encoding 0%
Landmark mapping (68 pts) 0%
Texture synthesis 0%
Lighting adaptation 0%
█ processing
✓ MODEL COMPILED — Face ready for deployment
OPERATION BRIEFING

Choose Your Weapon

Model compiled. Sarah Chen's face is ready to wear. Now choose the delivery mechanism for the live Zoom call with Finance Manager David Park. Goal: authorize a $4.2M wire transfer.

Open-Source Face Swap
DeepFaceLive · FOSS
90,000 GitHub stars
Free — no license, no API key
Consumer GPU — gaming PC as fraud machine
Real-time swap — 40ms latency on webcam feed
"Best for live calls. Artifacts imperceptible at Zoom's 720p compression."
Commercial Service
HeyGen · SaaS
$29/mo — no hardware needed
More polished — fewer visible artifacts
Pre-recorded mode only — limited live support
Cloud render — any laptop, no GPU
"Cleaner output, but real-time is limited. Requires a 'bad connection' cover story."
PROFESSOR PHISH NOTES

"DeepFaceLive processes your webcam and overlays Sarah Chen's synthetic face in real time. David Park will see her face. You speak. The algorithm stitches. At 720p over Zoom, the artifacts fall below most defenders' perceptual threshold."

"HeyGen is cleaner — but you lose real-time control. You'll need a pre-recorded clip and a 'bad connection' excuse. Riskier. Requires a solid social engineering script."

DEFENDER MODE — SPOT THE DIFFERENCE

This is exactly what David Park saw.

Drag the handle to compare side-by-side: the original video (left) versus the AI face-swap (right). Look carefully — try your best to notice any differences.

Modern deepfakes have surpassed human detection ability. MIT researchers found human accuracy at identifying AI face swaps is just 53% — barely better than a coin flip — even when people know they're looking for fakes.

AI face-swapped Zoom call
Original video recording
← ORIGINAL
DEEPFAKED →
◂▸
Drag handle left to reveal original ·
THE REAL TEST

Can You Tell?

Below are 6 faces. Some are real people. Some are AI-generated. Tag each one.

Results revealed. Scroll down for the most important lesson of this course.

Tag all 6 faces to continue ( remaining)
TRICK QUESTION

Every single face in this lab was AI-generated.

Sarah Chen — the CFO you watched — does not exist. David Park — does not exist. All 6 faces above — none of them exist.

Every face was generated in under 2 seconds by thispersondoesnotexist.com. No GPU rented. No money spent. Just a web request.

YOUR ACCURACY
tagged AI correctly
thought were real
detection rate
THE IMPLICATION

If you couldn't reliably distinguish AI-generated faces from real ones when you knew you were being tested, imagine trying to do it in a high-pressure Zoom call with a $4.2M decision on the line. Protocol is everything.

PHASE 4 — DEFENDER PROTOCOL

The 6-Second Check

Every employee on a finance team should run this check on any video call with a financial instruction. Six steps. Six seconds. $4.2M saved.

THE 6-SECOND CHECK — MANDATORY
$25B
deepfake fraud losses 2024
53%
human detection accuracy
Professor Phish
PROFESSOR PHISH — CLOSING

"You cannot outperform a neural network with your eyes. Stop trying. Build protocols, not instincts. Every wire transfer request that arrives over video call must have an independent verification path. No exceptions."

Change Alias

Choose your villain name, or roll the dice.

Share Feedback

Help us improve ScamAI University

Feedback received!

Thank you for helping us improve.