Binary Bot Detectors Misclassify ~39% of AI Browser Agents as Human; Two Behavioral Features Fix It
arXiv 2607.26935 (2026-07-29) argues the human-vs-bot label space is architecturally incapable of representing agent traffic: an MLP binary classifier misroutes 39.1% of real AI agent sessions as human and a SAINT transformer 34.5%, while adding an explicit third class yields agent F1 = 1.000 across all 30 runs. Against a five-level evasion ladder including GAN-generated trajectories and replayed human cursor data (2,299 evasion sessions), they observe zero agent misses in 22,990 per-seed predictions. An exhaustive search over 9,401 GBMs finds two features — mouse_event_rate and teleport_click_ratio — give 100% agent recall at every evasion level, because Playwright simply does not emit the raw pointer-move and wheel-delta streams a physical device produces.
↳ Follow the thread