Accuracy

What Prynt catches — and what it misses

Measured on real browsers on October 22, 2026, with exact 95 % confidence intervals. We publish the misses too: a number you cannot check is not worth much.

How these were measured

Real Chromium 156, Firefox 157 and WebKit 27.2, driven by Playwright against a private copy of the Prynt API. Results were read on the server, as a customer would. Intervals are Clopper–Pearson, which is honest with small samples — that is why 30/30 reads as "88–100 %", not "100 %". These are lab numbers: one machine, scripted sessions. Field numbers need real devices; we are collecting them with a consented panel and will publish them here.

Bots and automated browsers

Re-measured on November 1, 2026: 28 bot scenarios and 8 real-browser setups, 30 sessions each, on Chromium and Chrome 156, Firefox 159, WebKit 27.2 and Brave.

ScenarioFlagged95 % interval
No evasion: Playwright, Puppeteer, Selenium, undetected-chromedriver headless — every engine388/38899.1–100 %
One flag or one line: headed Chromium and Chrome with --disable-blink-features=AutomationControlled; Firefox with navigator.webdriver=false90/9096.0–100 %
Stealth tools: puppeteer-extra-stealth, playwright-extra, playwright-stealth, a hand-written stealth set, undetected-chromedriver238/26884.4–92.3 %
…of which undetected-chromedriver with a visible Chrome window0/300–11.6 %
Real Chrome driven by xdotool (no automation framework; keyboard and mouse events from the OS)0/600–6.0 %
AI agents that name themselves in the User-Agent (11 published tokens, simulated)110/11097–100 %
An AI agent that browses with a normal Chrome User-Agent and signs its requests (simulated)0/100–31 %

False positives on real browsers

Browser setupsFlagged95 % interval
Chrome, Chrome incognito, Firefox, Firefox private, resistFingerprinting, CanvasBlocker, Brave, Brave private — flagged as a bot0/2400–1.5 %
…flagged as tampering0/2400–1.5 %

What this means: Prynt catches automation frameworks as they come and with the usual one-line evasions, and about 9 in 10 sessions of the stealth tools, by finding the lies their patches leave (a proxied iframe, a page that claims a GPU while its Worker renders in software, emulated screens, hand-made plugin lists). It misses a real Chrome window driven by undetected-chromedriver and a real Chrome driven by OS-level input: there is no patch to find, and the open-source detectors we compared against miss them too. For those, use the behaviour signals and the JA4 TLS fingerprint at your edge, which scripts cannot patch from the page. An agent that presents itself as a normal Chrome is not identified as an AI agent. The browser setups above were lab browsers, not people; field numbers come from the consented panel.

Recognizing the same device

ScenarioResult95 % interval
Same device after clearing cookies and storage (each of Chromium, Firefox, WebKit)8/863–100 %
Same device in a private window (each engine)8/863–100 %
Firefox with resistFingerprinting: private window and cleared storage2/2 each16–100 %
Identical machines with different time zone, language and screen got different ids0 of 72 pairs—

What this means: clearing storage or opening a private window did not create a new identity. The other side of that: machines with identical hardware and browser — a fleet of the same corporate laptop, or look-alike phones — can share an id on their first visit. Collision rates on real, different devices are what the panel will measure.

The confidence score

Matches scored 0.9 or higher were the same device 99 % of the time in our benchmark. Lower scores overstate: a 0.3 does not mean "30 % likely". Treat anything below the match threshold as "not matched".

Related

Bot detection · AI agent detection · Edge & JA4 · Trust