A researcher planted two million canaries inside the AI supply chain and found that data does not stay private, according to an Ars Technica report published on 20 November 2026.
The report, written by Dan Goodin, describes a test built on prompts carrying unique, trackable secrets. Ars Technica's own description of the work states that the exercise proved data does not stay private, and the headline frames the outcome as a privacy disaster.
That is the limit of what is confirmed here. The available summary does not name the researcher, the AI companies or models involved, the number of planted secrets that surfaced outside their intended boundary, the rate at which they appeared, or how long the test ran.
What a canary is
In security research, a canary is a unique value planted deliberately so that its appearance anywhere else becomes evidence of leakage. Because the value has no other purpose, finding it somewhere unexpected is difficult to explain away. That is general background on the technique, not a confirmed detail of this particular test.
What remains unknown
The summary does not establish which layer of the AI supply chain was tested, whether model providers, data brokers or application developers were involved, whether any company responded to the findings, or whether the results were independently reviewed. The scale of two million planted markers suggests a broad sweep, but the published description gives no method and no measurement.
VANTIX Takeaway
AI products are often assembled from third-party models, data and tooling, which makes it harder to say where a user's input ends up. A headline alone cannot show how far that promise broke down in this case. Until the full findings are available, the report is best read as a signal that the question is being tested at scale, not as a measured account of the damage.
