A large-scale test of AI inference services found that sensitive prompts can escape through the chain of routers, resellers and model hosts that processes them. Security researcher HD Moore sent nearly two million prompts, each containing a unique bait secret such as an empty AWS account key, private repository reference or controlled URL.

The canaries exposed at least three compromised providers, resellers that logged prompts, and services that did not honor zero-data-retention settings. Attackers tested some AWS keys within hours and made tens of thousands of API calls, including attempts to inspect storage and create persistent access. Some requests had been stolen before a model ran, pointing to compromise at a provider’s front door.

Prompts are encrypted while moving between services but must be decrypted at each processing hop. A router may also pass a request to another router or one of several hosts, multiplying the parties able to log it. Moore recommends self-hosting for the strongest control; where that is impractical, buying directly from a first-tier provider reduces the chain. Privacy flags alone cannot protect data when an intermediary is already breached.