Bot detectors found between 9% and 99% of AI agent sessions

Proof of Human is a company that makes bot detection tools. It did a test of five bot detection systems against AI agents and automated browsers. The systems found between 9.1% and 99% of the automated sessions. Cloudflare Bot Management found 9.1%, and the system of Proof of Human found 99%. Each of five test agents did a task 20 times on each system. That gave 89 to 100 complete sessions for each system.

How it works

The test agents were scripted browsers, stealth browsers that conceal signs of automation, and AI agents. Each one did a task of five questions. The authors then counted how many sessions each system found as automated.

The numbers

• Proof of Human found 99% of all sessions.
• reCAPTCHA found 69.4%.
• Cloudflare Bot Management found 9.1%.

hCaptcha and Cloudflare Turnstile caught almost all scripted sessions but only some AI agent sessions. reCAPTCHA did the opposite.

What is not known

The authors tell that these systems are closed, and they cannot know which signals caused each result. The test had no human users, and the rate of incorrect alarms is not known. The authors tell that future tests must have more traffic and tasks that are not easy.

Insight: Sellers must now sort AI agents, not only bots

Forter tells that purchases by AI agents increased four times since August 2026. A report of 20 October 2026 tells that fraud tools frequently cannot find accounts that conceal their device. This test adds that most detectors find one type of agent and not other agents. Sellers must let some agents in and keep other agents out. It seems that the tool which reads behavior in a session will get the control.