Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.

by | Nov 21, 2026 | Technology

Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.

A security researcher conducted an experiment involving nearly 2 million AI prompts containing unique, trackable credentials to test how sensitive information flows through the artificial intelligence service supply chain. The study revealed significant vulnerabilities across multiple layers of the ecosystem, including AI routers, inference providers, and resellers that handle requests between end users and large language models.

The researcher embedded AWS access keys, passwords, private repository references, and other secrets into prompts routed through various AI service providers. Each secret was seeded into only one request to a single provider, allowing the researcher to track where data leaked and identify responsible parties. The findings exposed three separate hosting provider breaches, with stolen credentials used tens of thousands of times in AWS API calls within hours of being compromised. In some cases, attackers accessed stolen keys before models ever processed the requests, suggesting interception at the network’s entry points rather than deeper in the inference stack.

The investigation also identified intentional data harvesting by both resellers and model providers. Some resellers offered artificially low prices to attract prompts, then enabled logging to capture request data. Additionally, a Chinese AI laboratory’s systems fetched secret URLs from more than half of the test prompts sent to its models in the timeframe examined, including requests marked with zero data retention policies. Prompts routed through US hosting providers that resell this laboratory’s models also triggered the same data scraping, redirecting user data internationally despite customers’ expectations of domestic data handling.

The research underscores how plaintext decryption at each hop in the AI router network creates vulnerability windows. Although prompts are encrypted during transit, they must be decrypted at multiple intermediaries that determine routing and log activity. The study indicates that hosting models on premises remains the most effective safeguard against leakage, though cost constraints limit adoption of this approach.

Article Attribution | Read More at Article Source

Article summary produced by Claude AI