News

AI Supply Chain Test Finds Widespread Exposure of User Secrets

A researcher tested how well user data stays private across the AI supply chain by planting about two million unique, trackable “canaries” into prompts. These test secrets were designed to reveal where information travels and whether it can be exposed.

The experiment surfaced what the researcher described as a privacy disaster: data did not consistently remain confined to its intended context. The scale of the test indicates that leakage can affect a substantial volume of interactions rather than being limited to isolated cases.

The findings point to a broader security challenge for AI systems. Protecting user information requires safeguards across the full pipeline, including the services, tools, and workflows that process prompts. The article does not specify which particular components were responsible, but it makes clear that users cannot assume sensitive information will remain private merely because it is sent to an AI system.

Sources