Onyx Digital Intelligence.

The backlash to Google's selfie sign-in is aimed at the wrong thing.

Now those who know me and my work will immediately think I'm a hypocrite, I hope they stick around long enough to hear the argument.

On 23 July 2026, Google announced selfie video sign-in. Record a short video of your face, follow a few guided head movements, and that video becomes a way back into your account if you are locked out, forget your password, or do not have your usual phone or computer.

The replies arrived within the hour. Data harvesting. Digital ID. Bait. What next, DNA verification.

Most of it is wrong, and the parts that are right are right for the wrong reasons. That matters, because there is a real problem here and the noise is drowning it out.

First, the objection that does not work

The most upvoted complaint runs like this: my face is public, anyone can pull it off my Facebook, so my face is now my password and I have been walking around with it written on my forehead.

That objection fails, and it fails on the technology.

Two different things are happening when a system checks a face. Matching asks whether this face is the enrolled face. Liveness detection asks whether this is a real human physically present right now, rather than a photo, a screen, a mask, or a replayed clip. Blink, turn your head, open your mouth, plus passive signals like skin texture and micro-movement.

A photo pulled off a public profile should fail liveness. That is the entire point of the liveness layer. Holding a picture up to a camera is the oldest attack in the category and it is the one these systems are best at catching.

I do not like Google. I would take almost any excuse to roast them. On this specific point the comment section is wrong and Google is not.

What Google actually confirmed

Their own announcement and The Verge's coverage give four facts worth having.

The video is stored in the cloud. Not on the device. Google's wording is "encrypted at rest", which means Google holds the keys and can decrypt it. This is not end to end encryption and it is not the on-device model Apple uses for Face ID, where the template never leaves the secure enclave.

You do not need to take anyone's word for the cloud part. Google's own stated use case is getting back in when you "don't have access to your usual phone or computer". A template locked to a device you no longer hold is useless in exactly that scenario. It has to be server side for the feature to work as described.

It is a video, not a maths template. Most face systems store an irreversible embedding, a set of numbers that cannot be turned back into a face. Google is retaining the actual recording. If a password database leaks you change your password. You cannot change your face.

There is no infrared. This is the detail that matters most and it comes from The Verge, not from Google. Apple's Face ID projects thousands of infrared dots to build a depth map of your face, which is what makes a flat screen fail. Google's version works on any standard camera, phone or PC, with no added infrared light. No depth sensing. Liveness here is software inference on a flat 2D video stream. That is the weaker class of liveness, and it is consistent with Google describing the check only as "simple movements".

It is opt-in. Consent at setup, eligibility check, deletable at any time. Anyone arguing this was done to them without asking should stop. It was not.

What is missing from the documentation

Google's post runs about two minutes. There is no security whitepaper. There is no claim of certification against ISO 30107-3 (the presentation attack detection standard) or FIDO. There is no presentation attack testing data. There is nothing at all about injection attacks or camera attestation.

There is also a phrase that appears in both Google's post and The Verge's coverage, unexplained in either: your selfie video is used only for sign-in "unless you opt to share it for additional purposes".

Google never says what those purposes are.

That is not an accusation. It is an unanswered question about a video of your face sitting on someone else's servers, and it is the honest version of the data harvesting complaint that the replies were making badly.

Now the actual problem

Liveness detection defends against someone holding something up to a camera. It does very little against someone who never uses the camera at all.

A camera injection attack skips the lens entirely and feeds a synthetic video stream directly into the application, through virtual camera software or by hooking the device API. The app believes it is receiving a live feed from real hardware. Every movement prompt the system asks for can be rendered on demand.

Three pieces of evidence, stacked.

One: injection is trivial and already happening at scale.

When the UK's Online Safety Act age checks went live in July 2025, users began routing video game characters through virtual webcams to defeat the facial age estimation. Norman Reedus from Death Stranding 2 first, because the model is photorealistic and the photo mode lets you pose it to order. Then Garry's Mod ragdolls, which are not photorealistic at all.

Teenagers. A game they already owned. Free virtual camera software. Against a production system with regulatory force behind it.

An important limit on that story, because it cuts against the easy version: age estimation is not identity matching. Those checks ask one question, does this face look over 18. There is no reference image and nobody is checking whose face it is. A Garry's Mod ragdoll sails through precisely because nothing is being compared. That same ragdoll fails Google instantly, because it does not look like you.

What transfers is not the ragdoll. It is the pipe. The injection vector works, costs nothing, and is in use by people with no security background whatsoever.

Two: rendering a specific person's face in real time is solved.

I wrote about this on 28 May 2026, before any of this was announced. Incognia, a US fraud prevention company, demonstrated real time face swapping during a live video call. Not a pre-recorded deepfake. Live, moving, responding, hands in frame. Indistinguishable on screen.

That was the point of that piece and it stands: video is no longer evidence that the person on the other end is who they appear to be.

Three: the defence has a cost nobody can pay.

This is the number that should be leading every article about this feature and is leading none of them.

In December 2025, researchers at Verigram published an evaluation of virtual camera detection, the layer specifically designed to catch injection attacks. Over 30,000 real authentication sessions across Android, iOS, Windows, Linux and macOS. Their results table:

Read that again. To block virtually all injection attacks, you lock out nine in ten of your real users. The only deployable setting, the one where 85% of legitimate people get in on the first try, lets one in every ten injection attempts through.

Their own conclusion on the strictest setting: suitable "only for highly sensitive environments where security requirements outweigh operational accessibility". Forensic analysis. High assurance identity verification.

Not a consumer account recovery flow with billions of users and a phased rollout that Google wants people to adopt.

The same paper is explicit that liveness detection is a separate layer from virtual camera detection, and that liveness "may fall short when confronted with sophisticated video injection scenarios, where the spoofed content convincingly mimics live behavior".

Google documents the first layer. Google does not mention the second.

What this means for you, specifically

Nobody is assembling a real time face swap rig and an injection pipeline to get into my Gmail. The effort does not justify the return, and it will not for the overwhelming majority of people reading this.

That is the correct way to think about it. Security is not binary. It is cost against value. The right question is never "can this be defeated", because everything can be defeated. The question is at what cost, and against whom.

Two things follow.

If your account is worth real money or real leverage, a head of state, a crypto holder, an executive with signing authority, a journalist with sources, then the calculation is different and it was already different before this feature existed. Your face is on film. That is the raw material.

And attack cost only ever moves one direction. Today's bespoke rig is next year's subscription service. The Garry's Mod crowd are the proof of that: what was a research demonstration became a thing bored teenagers do, in about eighteen months.

What I am not claiming

Google has not said they lack injection defences. They said they use "standard security practices to detect and help prevent suspicious sign-in attempts", which is a catch-all broad enough to cover device fingerprinting, IP reputation, behavioural history and quite possibly injection detection too.

Undocumented is not the same as absent. Anyone telling you Google has no defence here is asserting something they cannot know.

And there is a real difference between Google's position and the UK age gate that got beaten by a ragdoll. The age check had no account context to work with, no history, no device, no baseline. It was a single anonymous question with nothing behind it. An attacker going after your Google account is walking into risk signals that have been accumulating for years. That is a genuine advantage and the criticism is weaker for ignoring it.

The questions worth asking

Not "how dare they". These:

That last one is the one I care about most. An account is only ever as strong as its weakest recovery route. It does not matter how good your front door is if there is a window.

Source ledger

Tier A, verified primary:

Tier B, credible but qualified:

Noted caveats:

Tier C, excluded:

Built under the #OnyxAudit method: Replicate. Document. Disclose. Punch at systems, not people. State facts, let readers conclude.

Adjacent to true is not true. The side of truth doesn't have a flag, only a standard.

Onyx Digital Intelligence @onyxaudit - 𝕏 / onyxdigitalintelligence85@protonmail.com