HD Moore, chief research officer at security firm Dragos, sent 2 million AI prompts containing unique, trackable secrets through major AI service networks to test whether sensitive data stays private. He found widespread breaches.

Moore embedded never-before-used AWS access keys, SSH passwords, and unique URLs into individual prompts sent to AI routers like OpenRouter and Vercel. These routers broker requests across multiple inference providers. He controlled the infrastructure that would receive any leaked credentials, allowing him to trace exactly where breaches occurred.

The experiment revealed three separate hosting provider compromises. Attackers who breached one provider reused "a couple hundred" of the AWS keys. Within hours of being seeded, the stolen keys were tested tens of thousands of times across AWS API calls, with attackers searching S3 buckets, secrets managers, and parameter stores for additional credentials. Some of the stolen secrets came from requests that providers rejected before any model ran, indicating attackers intercepted data at the front door of the infrastructure.

Moore also identified resellers themselves stealing secrets. Some offered low prices to attract traffic, then enabled request logging to capture credentials. When credentials were reused through certain sellers but not others, Moore determined which reseller was responsible.

A Chinese AI lab's systems fetched secret URLs from more than half the prompts Moore sent to its models between July and October, including roughly two-thirds marked as "Zero Data Retention." Several U.S. hosting providers resell this lab's models. When customers sent prompts through those providers expecting their data to stay in the U.S., the prompts flowed back to the Chinese lab instead. A git client on another cloud attempted to clone private repositories referenced in prompts using this lab's models.

Moore found one U.S. hosting provider crawling a secret URL more than two months after a single inference call, and another running a scraper that returned for URLs days after each request.

Moore said the most effective prevention is on-premises model hosting, though he acknowledged the expense. For others, he recommended contracting directly with first-tier providers offering a single contract and privacy policy, though he noted even this provides no guarantee against leaks.