A five-month investigation tested what happens to prompts after they are sent to AI services. From July through November 2026, researchers submitted nearly two million prompts to widely used providers, embedding unique URLs and credentials tied to controlled test systems and AWS accounts. These “canaries” were designed to be unusable in legitimate operations, so later attempts to access them could reveal that someone had encountered the prompt data. The central concern is that inference data may pass through providers, resellers, and other parts of the AI supply chain, where customers may have limited visibility into how it is handled.
The researchers report three kinds of exposure. First, activity involving the canary systems suggested that data from at least three AI providers had been compromised. Second, multiple resellers in an inference marketplace could view and retain customer prompts through ordinary logging, without exploiting a flaw or bypassing controls. Third, observed handling appeared inconsistent with privacy settings or published policies: prompts sent to U.S.-based providers were transmitted to Chinese AI organizations, unique links were accessed months after the original request, and one seeded link later appeared in an internal company chat in China. The findings frame canary tokens as a way to detect downstream prompt exposure that conventional security checks may miss.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.