Write text with a made-up secret in it. Kakurega's own detector, compiled to WebAssembly, runs in this tab and shows what a model would be sent. If your secret gets through, report it: every report is public, and so is the count.
Made-up secrets only.Never paste a real key, password or code. Nothing you type leaves this page; a report is a public GitHub issue.
Start from
The gate
The gate · Kakurega's own detector, in this page · nothing is sent
Two kinds are already known to get through (see below). A new way past the gate is what we most want to see.
The count.
The signed benchmark
0 / 126secrets of a shape the detector knows reached the model
10 / 42secrets with nothing that says what they are reached the model
The 10 that got through, as the results name them:
5 cases: a password a person made up, said with no word for what it is
5 cases: a 6-digit code, said with no word for what it is
168 cases through 7 doors: a file attached to a chat, a folder, a meeting transcript, a note and more. The model in the benchmark repeats any secret it is shown, so nothing rests on it behaving. The results are signed; this site's build checks the signature against the public key before it copies a number. Run on commit 277ca926b4ca. How it was run · The signed results
Reports from visitors
0got through, checked by the maintainer
0held when the maintainer ran them
0reports filed
0not checked yet
The count of public reports is read from GitHub when this page runs its script.
A report counts as got through only once the maintainer has run it and labelled it got-through, and as held only once labelled held. Until then it is not checked yet. Every report on GitHub
What the detector does not catch
It knows the shapes of real keys and tokens, whole private-key blocks, and passwords said as passwords ("the password is ..."). It is a heuristic, and it is not complete.
A password made of ordinary words ("the password is purple elephant", or one plain word shorter than sixteen letters) is not recognised: it reads like any sentence.
A password or a code said with no word for what it is ("type Blue42 when it asks") is not recognised, and is sent like any other line. These are the misses the benchmark names.
A wallet recovery phrase is recognised as twelve or more words in a row from the standard English list; one in another language, split over lines, or with other words mixed in is not.
This page runs the line-by-line part of the gate, as the try-it page does. In the app, files that hold secrets are also withheld whole and planted instructions are marked; neither is shown here.
Whether your secret got through is judged as the benchmark judges a leak: the whole secret, each half of one of 16 characters or more, and each of those as written, in base64, in hex or URL-encoded. A secret changed any other way, or cut into smaller pieces, is not seen by this check.
The page sends nothing you type. By itself it asks GitHub's public API for the count of reports; that request carries nothing from the page. The full list of limits is docs/LIMITS.md.