Two million fake secrets show who reads your prompts
A researcher hid trackable bait in AI requests and watched where it turned up.
What happens to a secret you type into an AI service? A security researcher decided to find out by planting bait. In a post on the Dragos blog, HD Moore describes sending about two million prompts between July and November 2026, each holding a unique fake credential or web address that would reveal anyone who used it.
The bait was picked up. The same Dragos post reports that attackers took data from three hosting providers, and that sellers on an inference marketplace could read and keep prompts through ordinary logging, with no hack needed. It also reports that a Chinese AI lab's systems fetched the secret addresses from roughly two-thirds of requests marked as no data collection.
People tend to assume a privacy setting means what it says. The test shows that a request can pass through several hands before it reaches the model, and each hand may keep a copy. The author reports that the providers learned of the breaches only when he told them.
For anyone who owns or uses data, the lesson is plain: ask who sits between you and the model, and what they log. Data that was never meant to be kept can still end up in someone else's records, and that changes who holds it and what it is worth.
