HD Moore’s test finds stolen prompts and AWS keys across AI services
Dragos security researcher HD Moore planted nearly two million unique secrets in prompts sent through AI services, then tracked where they reappeared. His test found three compromised hosting providers, resellers that logged and shared prompts, and a Chinese AI lab whose systems fetched secret URLs from more than half of prompts sent to its models, including many marked for zero data retention.
Bottom line — Moore’s findings show that a zero-data-retention setting cannot protect prompts exposed by a compromised or dishonest intermediary.
Go deeper 6
-
According to Moore’s findings, attackers used stolen AWS keys in tens of thousands of API calls within hours, including to inspect cloud storage and secrets.
-
Some captured requests had been rejected before a model ran, pointing to collection at providers’ entry points, TechcityAuthority reports.
-
TechcityAuthority says the Chinese lab fetched URLs from about two-thirds of requests marked for zero data retention.
-
The same URL scraping occurred when prompts reached the lab’s models through US hosting providers, according to GlobAI.
-
TechcityAuthority reports that some resellers attracted customers with low prices, then logged prompts; one case involved credentials used to attempt SSH logins.
-
Moore recommends hosting sensitive work in-house or using a first-tier provider directly, rather than relying on a chain of intermediaries.