Sciencepulse

Two Million Canaries Suggest AI Prompts Don't Stay Private

AI Privacy·November 21, 2026

Two Million Canaries Suggest AI Prompts Don't Stay Private

A researcher has tested how far private information travels once it enters an AI system, and the answer was unsettling. By planting roughly two million "canaries" into prompts, meaning unique strings that can be traced back to a single source, the researcher could follow what happened to the data afterward. The canaries turned up in places they were never meant to reach, which the researcher characterized as a privacy failure.

The method comes from security practice. A canary is a marker with no real value on its own, such as a fake account number or a unique token, that reveals exactly when and where it gets copied or exposed. Because each marker is distinct, a match points to the specific path it took. Using that technique at this scale gives a rare, concrete view of data flows that companies usually describe only in policy documents.

The finding matters because AI features rarely involve just one company. A prompt typed into an app may pass through the developer's servers, a model provider, plugins, logging services and analytics tools. Each of those hops is a place where data can be stored, reused or exposed. A user who assumes a conversation ends in one chat window may be wrong about how many parties touched it.

The available summary does not name the vendors involved or explain exactly how the canaries escaped, so it is hard to judge how severe the exposure was or whether it came from logging, third-party tools or model training. Those details will determine what companies need to fix. For now, the practical lesson is simple. Treat anything pasted into an AI tool as potentially retained, avoid sharing passwords, client records or other secrets, and check the data retention and sharing settings for the services you use.

Reporting based on an external source.