Behaviour check
Not only the answer: how it was given.
A script can read a captcha or guess a drag; it rarely behaves like a person while doing it. So the SDK notes how the person behaves while a challenge is on screen and sends it with the answer, and WAU Captcha checks both.
A hand and a script, moving the same distance
What the SDK records
- How long the challenge was on screen before the answer was sent.
- Where the pointer moved, relative to the widget, and whether it was a mouse, a finger or a pen.
- The time between key presses in the answer box, never the keys themselves.
- Whether the answer was typed, dictated or pasted, and events that a script dispatched instead of the person.
- Whether the browser says it is under automation, and how long the page was hidden.
- Three hints the browser gives about itself: a user agent naming HeadlessChrome, no preferred languages, a Chrome browser without window.chrome. A script can fake them, so they only lower the score.
- The answer to a small proof of work the browser solves for each challenge, in a fraction of a second: cheap for one visitor, costly for a script answering thousands.
What it never records
What the check decides
- Automated: events dispatched by a script, an answer typed into the box by a script, an answer faster than anyone could give it, a wrong proof of work, or a User-Agent of an automation tool or HTTP library. The answer is refused even when it is right.
- Suspicious: the answer counts, but the session looked scripted, for example a browser under automation, a mouse moving at one speed in straight lines, keys pressed at even intervals, no proof of work (an SDK before 1.3.0), or an answer from another address or browser than the one that fetched the challenge.
- Human: nothing out of the ordinary. Keyboard-only use, screen readers and dictation are ordinary.
What your server receives
When your backend redeems the token, /siteverify adds behavior: the verdict, a score from 0 to 1 and the reasons. Accept human and refuse not_measured; decide about suspicious per form, and refuse it for sign-up and login.
{ "success": true, "challenge_type": "image_distort", "challenge_ts": "2026-10-05T08:15:02Z", "behavior": { "verdict": "human", "score": 1, "reasons": [] }}Not measured
What we collect
The behaviour data is sent with the answer, judged in memory and dropped: it is never stored. What stays is the verdict, the score and the reason codes, in the usage history of the key, and with the token until the site's server redeems it. While a challenge can be answered, 2 minutes at most, Redis also holds the visitor's IP address, the website it was asked for from and a hash of the User-Agent, to check that the answer comes from the same browser. Redis also counts, per IP address, wrong answers for 10 minutes and answers judged automated for an hour, to slow down addresses that keep failing, and the automatic passes of auto_check for an hour, those without a clearance too, to cap them; and per auto_check clearance, its passes for an hour and a day.
Read the privacy policy