It started with a pair of Bluetooth headphones that refused to behave.
A developer was using multipoint headphones, the kind that hold a connection to two devices at once and hand audio back and forth between them. The handoff from laptop to phone stopped working. Nothing obvious explained it. Then they closed an AliExpress tab, and the headphones started behaving normally again.
That is the sort of coincidence most people shrug off. Instead, they went looking, and what they found on the AliExpress homepage was a pair of obfuscated scripts quietly building audio processing graphs through the browser and keeping a live connection open to the system’s sound hardware. No sound you could hear. No indicator anywhere. Muting the tab did not stop it.
The point of the silent audio was not to listen to anything. It was to identify the machine it was playing on.
The short version
- What was found: two obfuscated scripts on the AliExpress homepage creating running audio contexts wired to the system’s speakers
- How it was noticed: the audio contexts interfered with multipoint Bluetooth headphone switching. Closing the tab fixed it.
- What it does: the Web Audio API generates an inaudible signal, then measures how this particular device processed it
- Why that matters: tiny differences in CPU, audio stack, browser build and drivers produce a repeatable value that helps identify you
- Not a recording: your microphone is not involved, and nothing you say is captured
- What else was collected: canvas rendering output, display settings, hardware configuration and interaction signals, bundled and sent to Alibaba servers
- Who called it out: Brave, which says its browser has blocked audio fingerprinting by default for more than six years
Audio fingerprinting, explained without the jargon
Every browser ships an interface called the Web Audio API. It exists so that web apps can synthesize and process sound: a music sequencer, a game engine, a video editor in a tab. To do that, a page builds a small graph of nodes. Something generates a tone, something shapes it, something reads the result.
Fingerprinting abuses the last step. Set the volume to zero so nobody hears it, push a signal through the graph, then read the numbers that come out the other end. Those numbers are not identical across machines. Floating point math behaves slightly differently depending on the processor. Audio drivers round differently. Browser builds implement the same math with small variations. The output is stable on your device and subtly different on someone else’s.
On its own that value is not enough to identify anyone. Combined with a dozen other measurements, it gets close. This is the part people underestimate: no single signal is incriminating, and the combination usually is.
| Signal | What the site measures | Why it helps identify you |
|---|---|---|
| Audio | How a silent signal comes back out of the audio graph | Reflects your CPU, audio stack and browser build |
| Canvas | Pixel output from drawing hidden text and shapes | Reflects your GPU, drivers and installed fonts |
| WebGL | Renderer strings and how 3D scenes rasterize | Narrows you to a graphics chip and driver version |
| Display | Resolution, color depth, pixel ratio, available screen area | Unusual monitor setups are surprisingly rare |
| Hardware | Core count, memory, platform, touch support | Splits the population into small buckets |
| Behavior | Mouse movement, scroll cadence, typing rhythm | Separates humans from bots, and people from each other |
What actually happens in the two seconds after the page loads
The developer also documented scripts pulling canvas rendering results, display configuration, hardware details and user interaction patterns. Those were bundled with the audio value and shipped off to Alibaba infrastructure. Individually, unremarkable. Together, a profile.
Why a shopping site would bother
Here is the part that gets flattened in most coverage. Fingerprinting is not automatically an advertising play. It has a legitimate and very common security use, and marketplaces have a genuine problem to solve.
AliExpress deals with account takeovers, coupon farming, fake reviews, scripted checkout bots on limited stock, and refund fraud. Cookies are useless against all of that, because the attacker simply clears them. A device signal that survives a cleared cookie jar is exactly what an anti fraud team wants, and every large retailer, bank and ticketing site runs some version of it.
The problem is that the same measurement, taken by the same code, is equally good at building a persistent advertising identity that follows you across sessions and ignores every privacy control you thought you had turned on. The technique does not announce which purpose it is serving. You cannot tell from the outside, and neither can a regulator without seeing what happens to the data afterward.
The consent problem
Cookie banners exist because storing an identifier on your device generally requires asking first. Fingerprinting sidesteps that framing entirely, since nothing is stored on your machine at all. Regulators in the EU and UK have said for years that fingerprinting falls under the same rules as cookies, and the UK’s ICO has been explicit that it is harder to justify precisely because users cannot see it, cannot clear it and cannot opt out of it. Enforcement, so far, has not matched the rhetoric.
Brave’s response, and what it is actually claiming
Brave amplified the finding on X, writing that “Alibaba’s AliExpress was caught using users’ audio systems to track them. AliExpress wasn’t recording users but instead playing a silent sound and measuring how users’ specific devices processed it in order to fingerprint them.”
The second half of that sentence deserves highlighting, because plenty of people read the headline and assumed a shopping site was listening to their living room. It was not. No microphone permission was involved. Playback, not capture.
Brave says it has defended against audio fingerprinting by default for more than six years, using a technique it calls farbling: rather than blocking the API outright, it injects small amounts of randomized noise into the values a site can read. The site still gets an answer, so nothing breaks. The answer is just different for every site and resets between sessions, which makes it useless for stitching your activity together.
That approach is worth understanding because it differs from the alternatives. Firefox leans toward normalization, reporting the same generic values as everyone else so you blend into a crowd. Safari trims the surface area, returning simplified system information. Chrome, which still exposes canvas, WebGL, AudioContext and a long list of navigator properties, does the least of the four.
| Browser | Approach | Effort required from you |
|---|---|---|
| Brave | Randomizes canvas, WebGL and audio output per site and per session | None, it is on by default |
| Firefox | Normalizes signals, spoofs timezone and screen size, blocks canvas readback | On in private windows, opt in elsewhere |
| Safari | Reports simplified system information to shrink the fingerprint surface | None, but limited to Apple platforms |
| Chrome | No dedicated anti fingerprinting defense | Extensions and manual hardening, with mixed results |
| Tor Browser | Aggressive normalization so every user looks identical | None, but expect broken sites and slow pages |
One correction worth making, because it comes up every single time this subject does: private browsing does nothing here. Incognito clears cookies and history when you close the window. Your fingerprint gets computed fresh each time from hardware you did not change, so it comes out the same. If anything, the smaller population of people browsing privately makes you marginally easier to pick out.
What to actually do about it
You are not going to make yourself fingerprint proof, and anyone selling that idea is selling something. What you can do is stop being trivially linkable.
- Switch the browser you shop in. Not your whole life, just the tab where you compare prices on marketplaces. Brave or a hardened Firefox profile covers most of this.
- Turn on Firefox’s fingerprinting protection outside private windows if Firefox is your daily driver. It is a setting, not a download.
- Stop treating incognito as a privacy mode. It is a shared computer mode.
- Cut your extension count. An unusual combination of installed extensions is itself a strong identifying signal, which is the quiet irony of installing five privacy tools.
- Pick the app or the website, not both. Splitting your behavior across two identities gives any profile less to join up.
- Test yourself. The EFF’s Cover Your Tracks tool will tell you how unique your current setup looks in about thirty seconds.
The browser you choose matters more than any single toggle, which is one reason the fight over what browsers are allowed to block keeps mattering. When Chrome moved to finish off the old extension platform that real ad blockers depend on, it narrowed the options for anyone trying to defend themselves inside Chrome specifically. On phones the calculus is similar, and our ranking of Android browsers weighs tracking protection heavily for exactly this reason.
It also fits a pattern that has been hard to ignore this year: the surveillance most people should worry about is not dramatic, it is ambient. Sometimes it takes a vendor putting a warning directly on your lock screen to make it visible. Most of the time nobody tells you at all, and it takes a developer with misbehaving headphones and a free afternoon.
Neither AliExpress nor Alibaba has publicly addressed the finding. The scripts, as of this week, are still there.
