Your High-End Mouse Can Leak Speech—But Only Under Narrow Conditions
Mic-E-Mouse has not turned ordinary computer mice into a newly active mass-surveillance threat. The current public record still describes a proof-of-concept in which researchers recovered speech from vibrations captured by high-performance optical sensors. The revised Mic-E-Mouse paper, updated on November 17, 2025 and listed as appearing at ACSAC 2025, reports recognition results of roughly 42% to 61% on two speech datasets under controlled conditions.
That distinction matters: the work demonstrates a real side channel, not a microphone hidden inside every mouse. An attack needs a sufficiently sensitive device, a surface that carries speech vibrations, software able to collect high-rate movement data and conditions quiet enough to preserve a weak signal. As of August 2026, the cited public materials do not document widespread exploitation in the wild or a universal product recall.
What the researchers actually demonstrated
Speech creates pressure waves, and those waves can produce tiny mechanical movements in a desk. A high-performance optical mouse repeatedly photographs or otherwise senses the texture beneath it to calculate movement. When the surface vibrates, the resulting coordinate data can contain a faint imprint of that vibration even though the mouse has no conventional audio transducer.
The raw measurements are not ready-to-play recordings. Mouse packets arrive with irregular timing, limited resolution and substantial noise, while the sensor’s response varies across frequencies. The researchers therefore corrected the sampling, applied Wiener filtering and used a neural filtering system to enhance speech-related components. Their strongest controlled result was an improvement of up to 19 dB in signal-to-noise ratio; the reported recognition range depended on the dataset and experimental configuration.
This is best understood as a side-channel attack. The device performs its intended job—reporting motion—but its output unintentionally reveals information about a different physical process. The mouse does not suddenly gain microphone permissions, and the operating system does not label the collected coordinates as audio. That mismatch is the security lesson: protecting conventional microphones does not necessarily block information reconstructed from another sensor.
The software path is as important as the sensor
A remote attacker cannot listen merely because a vulnerable mouse is plugged in. Someone must collect the movement packets at a useful rate and transfer or process them. The researchers’ official project page describes a threat model involving compromised software—or potentially a web-based collection surface—and identifies games and other latency-sensitive applications as plausible carriers because frequent mouse input would be less conspicuous there. It also publishes the project’s code and experimental data.
“No microphone permission” therefore does not mean “no access requirement.” The attack still depends on code or web content obtaining enough sensor information, operating for long enough and delivering the measurements to a processing pipeline. Browser event-rate limits, application isolation, endpoint controls and restrictions on untrusted software can all affect whether that path is practical.
This also corrects a misleading shorthand often attached to the research: prior full-system malware infection is one possible route, but it is not the only collection scenario considered by the authors. Conversely, describing any ordinary webpage as capable of silently recording a complete conversation would go beyond the demonstrated results. The effective sampling rate available to the collector remains a decisive constraint.
Why the laboratory result does not transfer cleanly to every desk
The sensor must be unusually precise by everyday office-mouse standards. The experiments centered on high-performance devices with high DPI and polling rates, characteristics more commonly associated with gaming hardware. Greater sensitivity helps preserve small surface motions, but a high specification alone does not prove that a particular model will produce usable speech data.
The desk and the way the mouse is used matter just as much. Independent reporting on the experiment notes that the favorable setup used relatively thin, rigid surfaces, a mostly stationary mouse and a quiet environment; a rubber pad or mouse mat dampens the vibrations. The same analysis of the attack’s practical limits says the tested mice had at least 20,000 DPI and that real-world extraction would be challenging rather than comprehensive.
Normal use introduces another problem. Moving the mouse generates signals far larger than the subtle vibration caused by speech, while typing, footsteps, cooling fans and nearby conversations add competing energy. Filtering may recover fragments, but the published accuracy figures should not be read as a guarantee that an attacker can transcribe 61% of any conversation in an uncontrolled office. They describe particular tests on AudioMNIST and VCTK data, not a universal success rate.
What changed after the initial disclosure
The strongest confirmed update is bibliographic and technical rather than evidence of a growing campaign. The paper progressed from its September 2025 submission to a revised version in November and was identified with ACSAC 2025. The authors also made code and datasets publicly available, allowing other researchers to inspect the pipeline instead of relying only on a demonstration video.
No cited public page establishes that the technique has become commercial spyware, that all optical mice are affected, or that manufacturers have shipped a common fix. Claims that vendors are “working on” mitigations should not be treated as proof that specific firmware updates are available. Product-level guidance would require a model-by-model advisory that the current research record does not provide.
The absence of documented widespread use does not make the finding irrelevant. Sensor resolution and polling rates have risen because users want more responsive input, and those same improvements can increase the fidelity of unintended physical measurements. Mic-E-Mouse gives developers and hardware makers a concrete reason to minimize unnecessarily precise telemetry and to consider how quickly untrusted applications can read it.
Practical precautions without treating the mouse as a microphone
For most individuals, replacing a mouse is not the proportionate first response. The demonstrated attack already assumes a chain of favorable conditions, and breaking any important link reduces its usefulness. A soft or rubberized mouse mat is a simple physical barrier because it attenuates desk vibration before it reaches the sensor.
Software hygiene remains more important than unplugging peripherals. Avoid running untrusted executables, keep browsers and operating systems updated, review extensions, and restrict applications that have no credible reason to collect high-frequency pointer data. In managed environments, security teams can also watch for unusual outbound telemetry from games, creative tools or other applications that routinely process raw input.
People handling unusually sensitive spoken material can combine those measures with lower mouse polling rates or DPI settings where practical, although the published work does not establish a single safe threshold for every sensor and surface. The defensible conclusion is narrower than the alarming headline: a high-end mouse can carry recoverable traces of speech, but only when hardware, software and physical conditions align.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.