Researcher HD Moore conducted an experiment to test data leakage in AI supply chains by sending nearly two million prompts containing unique, trackable secrets. These 'canary' credentials, including AWS keys and private git repository references, were sent to various AI providers and routers.

The findings revealed that three hosting providers suffered security breaches where attackers stole hundreds of AWS keys. These stolen credentials were used tens of thousands of times to access AWS resources, including S3 buckets and parameter stores, often within hours of the initial prompt submission.

The study identified that some AI resellers log and store user prompts, potentially indefinitely, despite privacy policies. In some cases, resellers actively scrape secret URLs from prompts to gain unauthorized access to private data.

A specific Chinese AI lab was found to be scraping secret URLs from more than half of the prompts received between July and October. This activity occurred even for requests explicitly marked as 'no data collection' or 'Zero Data Retention'.

The research highlights that prompts are decrypted at every hop within the inference reseller network, making them vulnerable at each stage. Even when customers choose US-based providers to keep data local, prompts may still be routed to international labs that engage in unauthorized data collection.