Slider captcha: measure the gap instead of guessing

A slider captcha looks like the one challenge a language model cannot get wrong: drag the piece to the hole. In practice the model sees a picture, not a ruler, and a guess from a screenshot lands about 20 pixels off, which is enough to fail. Between 4 and 14 August 2026 our agent worked sliders on live widgets, and the runs that passed had one thing in common: the distance was measured before the drag.

What is a slider captcha?

A slider captcha is a challenge in which the visitor drags a piece or a handle until it fits a gap in an image, and the widget judges both the end position and the movement. GeeTest's slide puzzle is one example; its v4 documentation lists "Slide Puzzle" next to icon selection, match-three and Gobang, and describes the product as telling people from bots "through advanced behavioral analysis, environmental detection, and interactive challenges". Lemin asks for a piece to be fitted in two dimensions. The rotation variant replaces the drag with arrows that turn an image until it stands upright.

How does an agent measure the gap?

An agent measures the gap by comparing the shape of the loose piece with the image around it, and then drags the measured fraction of the width. Our agent uses three tools for it:

1. Take a screenshot of the puzzle element, so the image and the piece are in pixels the tools can read.
2. Call measure_gap with the image and the position of the loose piece. The piece is an element on the page, so its position is read from the page rather than guessed from pixels. The tool answers with the fraction of the image width the piece has to travel.
3. Drag with interact and that fraction, through the same input layer that moves every other pointer.

The comparison is why the piece is given to the tool rather than searched for. An earlier version searched the whole image for two matching windows, and GeeTest's circuit-board illustration repeats its own motifs: two background patches matched each other with a score of 0.91 and beat the real piece, and the drag landed 52 pixels off. A wrong answer with a high confidence is the worst kind, and the page already knew where the piece was.

Lemin needed one more step. Its widget swaps its elements faster than a reference to them survives, so the agent addressed the document container around the widget and aimed at a point inside it.

Scalebrowser gives each agent its own isolated browser with a persistent identity, on your own machine, so a run stays signed in, handles the captcha and finishes without anyone watching it.

See how Scalebrowser does it

Why does the rotation captcha reject every angle?

The rotation captcha rejects every angle because the demo compares the answer against a value its own arrows never produce, so no visitor can pass it through its interface. Our agent measured 15.3 degrees per click, turned the figure visibly upright, and tried 11 of 24 grid positions across the whole plausible range; every one was rejected, with the message "Incorrect captcha angle". During the check the page made no network request at all, so the rejection came from the page itself, which compared against a value outside the grid.

That result is one of the two open rows in our captcha solving record, and it stays open rather than being dropped. The measuring half works: measure_rotation read 14.8 degrees for one click with the wheel centred in the crop, against 15.3 degrees from an independent script on the same page. With the wheel off centre, the same click read 1.0 degree at a match score of 0.73, which is why the tool warns below 0.85. TikTok is a site that uses a rotation puzzle; in 3 logged-out loads on 11 October 2026 none appeared there, as TikTok's rotating puzzle records.

Does the drag path itself get scored?

Yes, the drag path is scored, and it can give a bot away even when the end position is right. In the BeCAPTCHA-Mouse study by Acien, Morales, Fierrez and Vera-Rodriguez, published on arXiv in 2020, a detector evaluated on 15,000 trajectories from 58 users and bots told realistic bot trajectories apart "with 93% of accuracy in average using only one mouse trajectory". The behavioral biometrics of a drag article shows how our own pointer compares with 120 people, measure by measure, and where it still differs.

What do you need to solve sliders yourself?

You need a way to read the distance and a pointer that moves like a hand. Read the piece's position from the page, measure the gap against its shape rather than its brightness, and drag the measured fraction in one continuous movement. Check the result by what the widget says, not by a text match: "Incorrect" contains "Correct", and a page that says the former passed a naive check for the latter in one of our runs. The verification page of the documentation lists the four measuring tools and the current status of every slider type.