The Viral Deep fake “Hack” That’s Actually Making You Easier to Scam

There is now an entire category of scams built around video calls. Billions of dollars have been stolen. This can be considered a new type of phishing scam.

Traditional phishing is when a bad actor impersonates someone trustworthy, a bank, a boss, a colleague, usually through email or text, to get you to hand over credentials or money. If you have a computer or a phone, you’ve definitely encountered at least one example of this before.

Video phishing calls raise the stakes. Many people have learned to be skeptical of links in emails. But a convincing face on a video call can easily feel like the real deal.

These attacks are already happening. Job seekers have been targeted by fake recruiters using deepfaked hiring managers to steal personal information. Employees have been on calls with people who look like their CFO or CEO only to later discover it was a deepfake used to authorize fraudulent money transfers.

Earlier this year, Google’s cyber security division Mandiant reported attacks by a North Korean group targeting cryptocurrency executives. The attackers compromised the Telegram accounts of trusted people in the crypto space, and used them to initiate video calls.

The calls were used as one step in a multi-stage attack that resulted in the victims installing malware.

They are high-value individuals in the tech industry, and they were still successfully manipulated. If you’re thinking “this would never happen to me” it's worth reconsidering.

All that is to say, this is the scenario you’re preparing for when you’re watching things like the ‘three finger trick’ video.

No. The difference between a deepfake and fully AI generated video is important, both for understanding how the trick works and for understanding why it sometimes won’t work.

Let’s start with the origin of the word itself. The term “deepfake” was coined on Reddit, and it’s a mashup of the words ‘deep learning’ and ‘fake.’

Deep learning refers to a subset of artificial intelligence (AI) where models learn patterns from large amounts of data. You can think of deep fakes as a far more advanced version of those 2016 era face swap Snapchat filters.

Deep fakes are built using neural networks, a core technique within AI, systems loosely modeled on how the brain processes information, trained on massive amounts of data until they learn to recognize and produce patterns. The specific architectures used in deep fakes include autoencoders and Generative Adversarial Networks (GANs).

An autoencoder is a type of neural network that learns to compress a face down to its essential features, the shape of the jaw, the distance between the eyes, etc. Once it understands those features it can reconstruct them and map them onto another face, altering perceived identity.

GANs, Generative Adversarial Networks, take this further. GANs include multiple neural networks running in competition with each other. One generates fake images. The other tries to detect them. They train each other through this back and forth: the generator gets better at faking, the detector gets better at spotting, and the output keeps improving. It’s an arms race baked into the architecture (which, as I’ll get to later, is also a useful way of thinking about the broader problem).

In contrast, accounts like clump.qc* create deep fakes. The base footage is real, but faces are swapped, often with recognizable public figures. One example included a deep fake of Ghislaine Maxwell appearing to be in Quebec, which spread widely before being debunked as a deep fake.

*This user’s account has been deleted since this video was posted. You can find it archived in several places here.

In short:

diffusion generated video = every pixel is AI generated
deep fake = faces mapped on top of real footage

Thanks for reading Technically, A Substack | By Mia! This post is public so feel free to share it.

Share

When you hold something in front of your face, fingers, a mug, whatever, it creates what’s called an occlusion. In the context of a deep fake, when your hand moves in front of your face, it blocks the view of key facial landmarks like eyes, nose, jawline, and mouth. The AI uses these landmarks as anchor points to track and map generated features onto the real footage underneath.

When those landmarks are obscured, the tracking breaks. It becomes unstable, and when it does, you can see it. The face glitches, the edges smear, and the image degrades into something that is not a real face.

These models are often built on older architectures that prioritized speed and price accessibility. They are doing a good amount of computational work in real time, and handling objects moving in front of faces is not something they were engineered for.

More sophisticated, resource-intensive deep fake systems already exist. And in these systems, the deep fake person can hold up three fingers, no problem. Occlusion doesn’t reliably break these models. They use 3D modeling and temporal data to infer what a face looks like even when covered. UC Berkeley deep fake detection specialist Hany Farid has been warning about exactly this: the next generation of systems has already solved the issues people are learning to associate with deep fakes.

No… not yet.

For the category of deep fakes that people are most likely to encounter, the trick can still work. Use it. But DO NOT let it become the whole of your understanding.

A scammer doesn’t need a perfect deep fake, just one that passes your one “quick and easy” detection trick. That is a low bar.

This is true beyond deep fakes. Any time you reduce verification to a single point of failure, you’re actively making yourself easier to fool.

And there’s a broader dynamic at work here, that’s known in cybersecurity as an adversarial system. This term is used to describe any environment where two opposing sides are adapting to each other in real time, each evolving in direct response to what the other does.

This is what I want you to take from this article, and the Huntress Labs clip: not the ‘three finger trick’, but why the trick works.

If you understand why the trick works and when it won’t, then you also understand how deep fakes work. And then you can come up with your own tricks that adapt to different scenarios. For example:

No one trick is foolproof. Think of them not as a single test, but as tools in an investigative tool box. And a toolbox can adapt as the tech evolves. A single trick will not adapt, and even worse, it will leave you vulnerable.

More importantly, these deep fake detection “tricks” are the last line of defense in a phishing zoom call scenario. There are earlier lines: skepticism about unexpected meeting requests, caution around links arriving through communication channels that feel off, awareness that attacks are becoming trickier.