A five-month investigation traced what happens to AI prompts after they are sent to popular services. From July through November 2026, researchers submitted nearly two million prompts, each containing unique URLs or credentials tied to systems they controlled. These canary tokens were designed never to be accessed in normal use, so any later attempt to reach them could reveal that someone in the AI supply chain had seen or retained the prompt. Monitoring the tokens uncovered signs that data from at least three AI providers had been exposed: seeded credentials were used in attempts to access test systems and cloud accounts.
The investigation also found that inference-marketplace resellers could inspect and keep customer prompts through ordinary logging features, without exploiting a software flaw or evading platform safeguards. In other cases, providers appeared to disregard privacy settings or stated data-handling rules. Prompts sent to U.S.-based services were reportedly routed to Chinese AI organizations; some unique links were accessed months after the original request, and one appeared in an internal company chat in China. Together, the findings show that prompt confidentiality depends on more than the AI model or direct provider: resellers, routing partners, logging systems, and downstream data practices can all expose sensitive information.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.