Spotting Canary Strings and Honeypot Files Before You Touch Them

What a canary actually is

A canary is a marker: a string, file, or record that has been placed somewhere it has no business being, so that its appearance elsewhere proves that someone or something crossed a boundary. The benchmark variety is embedded in training corpora — a unique phrase like a nonsense surname — so that if the phrase turns up verbatim in a model's output, the benchmark items it sat in were in the training data. The honeypot variety is placed in reachable locations: a credential file in a repo, a 'backup.zip' on a share, a page that looks more interesting than its neighbours. The principle is the same in both cases: the marker must be unique enough to be unmistakable and quiet enough not to attract attention except from the behaviour it is meant to catch.

Two properties make canaries work. Uniqueness, so a hit cannot be coincidence — the string must never plausibly occur otherwise. And quietness: the bait sits among ordinary data, indistinguishable from it until someone correlates events after the fact. A good canary file in a repository looks like a misconfigured config file with a plausible name; a good benchmark canary reads like any other line of synthetic training text.

The third property is the one people miss: the marker does not need to be invisible to a determined reader. Defenders only need it to be unique and quiet, because the record of the