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

What happened
Researcher HD Moore conducted an experiment by embedding 2 million unique secrets into AI prompts sent through brokers like OpenRouter and Vercel. These prompts were routed across various AI providers to test how well sensitive data was protected. The experiment uncovered that many secrets were leaked, including through compromised providers and resellers who logged and misused the prompts. Some providers also ignored privacy settings, allowing further exposure of sensitive information.
Key details
- 2 million unique secrets were used as digital markers in AI prompts.
- At least three AI providers were found to be compromised by attackers.
- Some resellers logged prompts and reused credentials for unauthorized access.
- Secrets marked with privacy flags like Zero Data Retention were still leaked.
- A Chinese AI lab scraped secret URLs from over half the prompts it received, including those routed through US providers.
Why it matters
Developers and businesses using AI brokers and providers face a significant risk of sensitive data exposure, even when privacy settings are enabled. Customers who rely on US-based providers to keep data within the country may have their information sent abroad without their knowledge. The findings mean organizations must reassess how they handle confidential information when using third-party AI services, as current safeguards may not prevent leaks or misuse.
