What we know
A researcher inserted 2 million unique, trackable "canaries"—secret pieces of data—into the AI supply chain to test whether sensitive information remains private. These canaries later appeared in the outputs of AI models, showing that user data is leaking at a concerning rate. This finding highlights a significant privacy vulnerability in how AI systems handle and protect data.
Why it matters
Embedding unique, traceable secrets into datasets is a method used to detect unauthorized data exposure. The discovery that these canaries surfaced in AI-generated outputs suggests that private information can escape containment within AI models. This exposes risks for individuals whose data may be unintentionally revealed and raises questions about the effectiveness of current AI data protection measures.
What is still unknown
No independent verification of these findings has been reported. Details about which AI models or languages are affected, the scope of the data leakage, and any planned fixes or timelines for addressing the issue remain unclear.

