Researchers HD Moore inserted two million unique digital canaries into the global AI supply chain to test for data leakage. These were valid AWS access keys, SSH passwords, and private Git repository references embedded in single prompts sent through various brokers like OpenRouter and Vercel. The experiment aimed to see if these sensitive secrets would appear in responses from the underlying models or be logged by the routing services.
The results confirmed that the ecosystem acts as a sieve for user data. Many of the canaries were recovered from model outputs or logs, proving that sensitive information entrusted to AI services is frequently exposed. This discovery highlights a critical vulnerability in how developers rely on third-party routers to find the cheapest or most suitable inference engine. Users cannot assume their prompts remain private when routed through these aggregators.
* The test involved exactly one secret per provider to pinpoint the leak source.
* Recovered data included credentials for empty AWS accounts and private code repositories.
* The exposure rate was found to be alarming across the tested infrastructure.




