AI / Published story
Researchers hid keys in 2 million AI prompts and watched them get stolen. OpenRouter has now cut three providers.
A Dragos experiment seeded about two million AI prompts with unique bait keys and links. Some were stolen from three breached AI hosting providers; others were taken by marketplace resellers or fetched despite “zero data retention” settings. OpenRouter says it has off-boarded three providers and removed 125 endpoints from its zero-retention list.
Our view
The lesson is not that AI services are unusable. It is that a prompt is a document handed to a chain of companies, and the cheapest chain is often the one you know least about. Zero Data Retention is a label, and this research shows labels can be wrong. OpenRouter’s response points the right way, but continuous testing stays a promise until outsiders can check it. Until then, treat anything in a prompt as something a stranger may read, and keep keys and passwords out of it. This is SIGNAL’s reading of Dragos’s report and the providers’ notices.
