Two Million Canary Tokens Reveal Critical Privacy Gaps in AI Supply Chain

Two Million Canary Tokens Reveal Critical Privacy Gaps in AI Supply Chain

Researchers have uncovered a widespread failure in data security across the artificial intelligence supply chain by deploying two million digital tripwires, known as canaries, to track the flow of sensitive information. The experiment highlighted how easily user data can be exposed as prompts travel through increasingly complex networks of AI service brokers.

The study was designed to investigate whether confidential information leaks when requests are routed through ‘AI gateways.’ These intermediaries, such as OpenRouter or Vercel, handle massive volumes of prompts from end users and direct them to various inference engines based on pricing or suitability. By using these routers, developers often bypass direct connections with major providers like OpenAI or Anthropic, creating a longer chain of custody for sensitive data.

HD Moore, the lead researcher, seeded specific, unused secrets into individual prompts sent to different providers to monitor for leaks. These canaries included unique identifiers such as AWS access keys tied to empty accounts, SSH passwords, private Git repository references, and custom DNS names. Because Moore controlled the infrastructure receiving these credentials, any unauthorized access would immediately identify the specific link in the chain responsible for the breach.

The results indicate that the AI ecosystem currently leaks highly sensitive data with alarming frequency, posing significant risks regarding security breaches, prompt theft, and privacy violations for users relying on third-party routing services.

3 responses to “Two Million Canary Tokens Reveal Critical Privacy Gaps in AI Supply Chain”

Leave a Reply

Your email address will not be published. Required fields are marked *