Researcher HD Moore conducted an experiment planting 2 million unique "canaries"—including AWS access keys, SSH passwords, and private Git repository references—into prompts sent through various AI services. According to Ars Technica, these digital tripwires were used to identify where sensitive information leaked within the infrastructure of AI routers and gateways that act as brokers between applications and model providers.
The study found that sensitive data sent through these intermediaries can be exposed to third parties as prompts are routed to different inference engines to optimize for cost or performance. To identify leaks, Moore seeded each secret into a single prompt and monitored for unauthorized access attempts using those specific credentials. Organizations using AI routing services can detect similar exposures by monitoring for any activity on unique, otherwise unused credentials seeded into their prompt streams.

