Firefox 143 cuts modeled unique browsers from 75% to 46% on ~25 million users

Firefox 143 shipped data-driven fingerprinting protections built from telemetry on almost 25 million Firefox release users. Collection used Oblivious HTTP so the study could not identify or track individuals. Researchers at Mozilla, the University of Toronto, and UC San Diego modeled three commercial fingerprinters and reported that stacking the shipped protections cut users those models treat as unique from 75% to 46%, weighted across four operating systems. The paper, appendix, and aggregate data are published at fingerprinting.study. Uniqueness is not the sum of fifty weak probes. Among users the Fingerprint Pro metric set marked unique, 80% were already unique after 8 measurements, out of more than 50 collected. Signifyd reached 80% of its unique set at 9 measurements while collecting more than 80. The distribution follows a rough 80/20 split: a short list of high-impact attributes, in combination, does the work. The paper also warns that a single headline percentage is suspect, because Windows, macOS, Linux, and Android do not share the same fonts, GPUs, or screen sizes. Two canonical priors bound the claim. Eckersley (PETS 2010) measured 470,161 self-selected Panopticlick visitors and found 83.6% unique. Gómez-Boix, Laperdrix, and Baudry (WWW 2018) measured about 2 million browsers and found 33.6% unique, concluding most clients hide in common configurations. The Firefox figure is not a rebuttal of either study. It scores commercial feature sets on an unbiased release-channel Firefox population, not a volunteer privacy audience and not the full multi-browser web. Treating 46% as proof those users are re-identified across sites overreads a model. For daily use the residual matters more than the delta. A drop of 29 points leaves 46% of the measured population unique under the strictest model. Tor Browser's resistFingerprinting clamps more APIs and accepts more breakage; Firefox 143 kept the compatible subset. Individuals reduce exposure by staying on default fonts, window sizes, and language lists and by skipping rare extensions, the surfaces that dominate the first eight metrics, rather than blocking every low-entropy API. Later movement off 46% is checkable in the public aggregate tables, and only with the same per-OS split.