Security
Massive Privacy Breach Found in AI Supply Chain Through Researcher's Canary Experiment
A security researcher planted millions of tracking signals through AI services and uncovered widespread leakage of sensitive user data across the entire ecosystem.
Ars Technica
Researchers have discovered a severe privacy crisis in the AI supply chain after deliberately seeding millions of hidden tracking signals—called "canaries"—through a network of widely-used AI intermediary services. The experiment revealed that sensitive information flows through these systems with alarming ease, exposing credentials and private data at scale.
How the Experiment Worked#
Security researcher HD Moore designed an ingenious test to expose data leakage across the AI broker ecosystem. He embedded 2 million unique digital trackers—AWS access keys, passwords, SSH credentials, private URLs, and DNS entries—each placed into individual prompts sent to different AI routing services. Because Moore controlled the infrastructure receiving any request using these credentials, he could immediately detect when and where secrets were exposed, and trace the exact path the data took through the system.
The AI Router Problem#
Many developers building AI applications don't directly use major providers like OpenAI or Anthropic. Instead, they rely on intermediaries called AI routers or AI gateways—services such as OpenRouter and Vercel that act as brokers. These platforms receive millions of prompts from end users and distribute them across available inference engines based on cost, capability, or other factors. This added layer was meant to simplify development, but Moore's research suggests it's become a major privacy vulnerability.
What Was Exposed#
The findings point to a system that leaks sensitive data extensively. The specific details of what was compromised and how widely the exposure spread remain to be fully detailed, but the initial results indicate that credentials, private URLs, and other secrets routinely escape containment as they pass through the AI supply chain, potentially reaching unintended third parties.

