Cloudflare Turnstile misses all AI-agent browsing, Proof of Human finds

The test set out to compare how well five bot-detection systems recognize two different things: scripted bots following fixed steps and AI agents that read a page and decide their next action. The scripted bots came in two forms — Chromium, which leaves automation traces, and CloakBrowser, which hides them and was configured to spoof browser data and use residential IP addresses. The gap between the two categories is the central finding: hCaptcha and Cloudflare Turnstile stopped the fixed-step bots at 95–100% but let most AI agents through. The researchers attribute part of this to how each check works. hCaptcha's image-selection puzzles tripped up the scripted bots but were solved by the three AI agents, which could recognize and answer them. Cloudflare Turnstile inspects browser state and, according to the team, let all three AI agents pass without ever presenting a challenge. reCAPTCHA v3, which scores users instead of posing puzzles, detected most agent activity. Yet even it caught none of the CloakBrowser runs. The team said it cannot say from this test whether the spoofed browser information, the residential IP address, or both were responsible. One limit remains open: the study did not measure false positives, and the researchers say human participants should be added to assess how often the systems mistake people for bots.