Resemble AI says bot tools missed most disguised AI agents

A script of about 25 KB watches scrolling, typing rhythm, clicks and the mouse pointer’s path, and Resemble’s servers do the judging. The benchmark sent more than 2,000 agent visits, from ChatGPT, Claude and Meta’s Muse to stealth browsers (browsers altered to pass as a person’s), to sites built on 12 different web technologies. Against agents that make no effort to hide, all seven tools scored from 94.8 per cent to 99.3 per cent or more. The catch is false alarms: at its default setting RAD-S1 wrongly flags about 5 to 6 per cent of real people as agents, Resemble says. PPC Land points out that the company ran the benchmark itself, with no outside check.

AI images can carry a signature: our watermark data shows ChatGPT and Nano Banana image files carry C2PA labels (a record of origin stored in the file) and SynthID marks (Google’s invisible watermark), checked 12 October 2026. For site visits, the nearest equivalent Resemble lists is Web Bot Auth, a signature sent only by agents that identify themselves.

At the default setting RAD-S1 marks roughly one human visit in every 17 to 20 as an agent, so a shop that blocked on its verdict would turn away human visitors at that rate; Resemble’s own advice is to label first, and its tool blocks no one by itself.