GPU Render-Timing as an AI-Resistant CAPTCHA: Software-Rendered Bots Separate From Real GPUs by ~5x

arXiv 2607.23389·medium signal

Rather than posing a puzzle AI can solve, this pilot (arXiv 2607.23389) measures the physical timing behavior of a client's GPU under a controlled WebGL workload — classifying rather than identifying, so unlike WebGL fingerprinting it leaks no persistent identifier. A 12-hour passive deployment characterized the in-the-wild adversary: 207 unsolicited requests, 86% automated, and 85% of browser-claiming clients failing HTTP header-consistency checks. Software-rendered automation, the dominant real-world adversary, separates from genuine GPUs by roughly 5x in mean render time; in a confound-controlled comparison holding GPU family and browser engine fixed, headless automation on real hardware still separates from human samples by 75–106% on frame jitter, timer-quantization ratio, and coefficient of variation. The authors explicitly scope this as pilot-scale on a single GPU architecture.