Bot detection decided on the server — and measured
Catch headless browsers, automation frameworks, scrapers and AI agents — decided server-side from raw tells, TLS fingerprints and how a real hand actually moves.
Why client-side bot checks fail
Anything the browser computes, a bot can patch. A script that returns isBot: false
is trivially overwritten. Robust bot detection must be decided server-side from
raw evidence the automation can’t consistently fake. Prynt’s agent ships tells — not a
boolean — and the engine weighs them where the bot can’t reach.
The tells that give automation away
- Headless & automation flags:
navigator.webdriver, Selenium/Puppeteer/ Playwright globals, CDP artifacts, missingwindow.chrome, empty plugin/language lists. - Native-function tampering: patched
getContext/toDataURL/getParameter(canvas-noise and UA-spoof tools betray themselves here). - SwiftShader / software GL under a non-headless UA, iframe hardware mismatches, and fresh-iframe webdriver re-reads that bypass a top-level override.
- JA4 TLS fingerprint captured at the edge, before any JS runs — a browser UA riding a curl/Go/Python TLS stack is a spoofed client.
- Behavioral biometrics: teleporting cursors, zero-delay typing and a motion sensor that never moves are flagged. Scripts that type and move at a human-like pace are not caught by this signal alone — see the measurements.
AI agents are the new bot surface
Agentic-AI traffic is the fastest-growing traffic class of 2026. Agents that name themselves (GPTBot, ChatGPT-User, PerplexityBot, ClaudeBot and other agents that name themselves) get their own verdict; agents that pass as a normal Chrome do not, yet. Prynt classifies it as a distinct verdict (not just “bot”) so you can allow, throttle or block AI agents on your own terms.
What it catches, and what it does not
We measured it on real browsers: automation frameworks are caught every time, with or without the usual one-line evasions, and stealth plugins about 9 times in 10 (238 of 268 sessions), with no real-browser setup flagged (0 of 240). What still gets through is a real, visible Chrome driven by undetected-chromedriver or by OS-level input, where there is no patch to find. The numbers, with their intervals and method, are on the accuracy page. That is why the decision also uses the network layer (IP, JA4) and why you can wire it to a challenge instead of a hard block.
Detect, then act — at the edge
Detection without mitigation is half a product. Prynt wires the verdict to Prynt Challenge — a risk-adaptive proof-of-work (a Turnstile/Arkose alternative) — and to Prynt Edge connectors (Cloudflare Worker, nginx/OpenResty) that allow, challenge or block at the edge, before traffic reaches your origin.
Related
Device fingerprinting · Account-takeover prevention · Prynt vs FingerprintJS · Glossary