AliExpress quietly ran two WebAudio graphs on its shopping homepage. They generated a waveform, measured the browser’s response, reduced the output to silence, and connected the graph to the system audio destination. A researcher found them because his Bluetooth headphones stopped playing audio from his phone whenever the AliExpress tab stayed open.[1]
This is a revisit to Chrome 148 Made Browser Math a Fingerprint, which covered operating-system math differences leaking identity through ordinary web APIs. The new development is a live merchant deployment with an observable hardware consequence: an obfuscated fingerprinting bundle exercised WebAudio strongly enough to seize priority on a multipoint Bluetooth path while remaining invisible to the page’s normal media controls.
The fingerprint itself appears weak in current browsers. The permission failure is excellent. A page performed hidden audio computation, touched the system output route, interfered with another device, and gave the user no useful indicator or switch.
The headphones became the debugger
Matthew Callaghan normally listened to music from his phone through multipoint headphones while keeping a Windows PC connected for notifications and occasional video. Opening AliExpress in Firefox or Chrome interrupted the phone. Closing the tab restored it. Muting the tab, browser, or Windows did nothing.[1]
The usual media suspects were absent. Callaghan found no audio or video elements, media playback calls, active Media Session, or media requests. The interruption began several seconds after page load. Instrumenting AudioContext and AudioNode.connect() exposed two running contexts created by heavily obfuscated Alibaba scripts named collina.js and fireyejs.js.[1]
Both scripts lived under an AWSC asset path associated with browser security and anti-abuse tooling. Their graph followed the same basic route.[1]
Zero gain kept the waveform inaudible. Connecting the graph to AudioContext.destination still made the browser process it as live output. That distinction explains why tab mute failed. The browser had no conventional media element to stop, while the operating system still saw an active audio route.[1]
This is the useful weirdness in the story. The feature announced itself through a physical side effect that its designers had treated as irrelevant. Obfuscation hid the code from ordinary inspection. Bluetooth multipoint converted hidden state into a blunt interface event: the music stopped.
One stale probe inside a much larger fingerprint
The audio graph measured how the browser processed a known sawtooth waveform. Audio implementations can produce slightly different results because of browser code, math libraries, CPU instructions, architecture, and hardware. Combined with other signals, those differences have historically contributed to browser fingerprints.[1][2]
AliExpress’s bundle gathered far more than audio. Callaghan found canvas rendering, WebGL details and shader precision, screen dimensions, device-pixel ratio, CPU and memory hints, installed plugins, supported media formats, WebRTC behavior, performance timing, interaction events, motion and orientation data, and properties associated with browser automation. The scripts serialized and encrypted results, then sent data to Alibaba telemetry services.[1]
The client code proves collection and transmission. It does not reveal retention periods, server-side linkage, or whether the result becomes a persistent identity, fraud score, bot score, coupon-abuse signal, or some mixture of them.[1] AliExpress has obvious reasons to separate ordinary shoppers from scraping, account takeover, fake accounts, payment fraud, review manipulation, and automated promotion abuse. Those reasons do not explain why broad device measurement runs on the general storefront before a sensitive action.
The WebAudio component may be fossil code. Firefox engineer Tom Ritter tested the extracted fingerprint and described it as nearly useless against modern Firefox. Mozilla made WebAudio output substantially constant in Firefox 118. In its telemetry, 99.24 percent of users landed in three values, with the remaining large buckets explained by CPU instruction differences. Mozilla has already collapsed two of those buckets and tracks work on the last major split.[2]
Ars Technica reported that Google considers the technique ineffective in Chrome because Chrome also ships controlled audio libraries. Safari likely has similar defenses, although Apple did not confirm that point for the report.[3]
A weak probe can survive inside a commercial anti-abuse bundle for years because the bundle rewards accumulation. Each signal looks cheap. Nobody owns the job of deleting the stale one. The result becomes a dense private questionnaire addressed to the browser, asked on every visit, with no public accounting of which answers still carry weight.
Browser permissions describe nouns, not computation
Browsers expose clear permission rituals for microphones, cameras, location, notifications, MIDI devices, and screen capture. Those prompts attach to recognizable resources. WebAudio usually enters through a different door because synthesis and processing are legitimate page behavior. Games, instruments, editors, accessibility tools, visualizers, and music software need low-friction audio graphs.
The AliExpress case breaks the comfortable taxonomy. The page did not capture microphone input. It created output computation. The output was silent. The computation still affected hardware routing. Existing indicators centered on audible playback and media elements gave the user a false picture of the page’s capability state.
The Hacker News discussion moved quickly toward speaker permissions and per-site audio controls. Callaghan clarified that conventional tab muting did not stop the behavior because the graph bypassed the media path users expect the mute control to govern.[4] A generic audio permission prompt would create fresh garbage, since users would approve it for product videos and every web instrument would inherit another click wall. The useful fix needs finer semantics.
Browsers can distinguish at least four states: audible media playback, silent processing disconnected from hardware output, active WebAudio connected to an output device, and input capture. User controls should expose the state that crosses into the hardware route. A site that connects a live graph to the destination while producing zero audible energy deserves an indicator and a working kill switch. Persistent background use deserves stricter treatment than a graph created after a click inside a music tool.
Anti-fraud code has become private browser policy
AliExpress appears to have packaged these measurements inside security infrastructure rather than an ordinary advertising script.[1] That placement matters. Anti-fraud systems receive broad tolerance from companies because blocking bots, stolen accounts, payment abuse, and promotion farming carries obvious economic value. Their false positives and privacy costs remain hidden inside risk scores.
The browser becomes an involuntary sensor array for that scoring system. Canvas answers one question. WebGL answers another. Timing, motion, audio, plugins, automation markers, and interaction cadence fill in the rest. The merchant decides which combination counts as a human, a repeat customer, a suspicious device, or a promotion abuser. The user receives a CAPTCHA, checkout failure, account review, or nothing at all.
Callaghan blocked the two observed script families with narrow uBlock Origin rules and reported that ordinary browsing continued without the hidden audio contexts. He warned that login or checkout could react differently because the scripts appear tied to anti-fraud controls.[1] That trade exposes the control surface cleanly. Refusing the device probe may change how the marketplace treats the device.
The audio fingerprint deserved ridicule because modern browsers drained most of its entropy. Its hardware side effect deserves attention because it revealed the machinery around it. One shopping tab reached through a silent processing graph, changed an external device’s behavior, ignored the visible mute model, and delivered a broad encrypted fingerprint to a private risk system.
The headphones did the one thing the interface refused to do. They told the user the page was active.
Sources
[1] https://blog.laserphile.com/2026/08/aliexpress-webpage-keeping-multipoint.html [2] https://ritter.vg/blog-webaudio_alibaba.html [3] https://arstechnica.com/security/2026/08/aliexpress-caught-fingerprinting-visitors-after-sending-inaudible-sounds-to-browsers/ [4] https://news.ycombinator.com/item?id=49372583