In July 2026, the Dragos research team set out to answer a question most AI users never ask: what actually happens to your prompts after you send them to an AI service? Over five months, they injected almost two million unique prompts—each with a trackable canary token—into major AI providers and watched where those tokens surfaced. What they found was not just theoretical: data was stolen, privacy settings were ignored, and breaches went unnoticed by the affected companies.
If you assumed that setting strict privacy options or using a reputable provider would keep your prompts secure, these findings suggest otherwise. The reality of the AI inference ecosystem is that prompts can pass through multiple routers, marketplaces, and resellers—each with their own logging, retention, and, sometimes, security blind spots.
How canary tokens traced the leaks
Dragos generated millions of unique URLs and credentials—canary tokens—and embedded them in AI prompts. Since these tokens were never used for any legitimate purpose, any access attempt would mean someone had seen the original prompt. This approach allowed the team to follow the journey of inference data through the supply chain, catching both accidental and deliberate leaks.
Compromised providers and marketplace resellers
The investigation uncovered three separate AI hosting providers where attackers had direct access to customer prompts. Dragos saw their seeded AWS credentials being used in attempts to access controlled test systems. In one breach, every request sent during certain windows—hundreds of unique AWS keys—was stolen. The companies affected had no idea until notified.
Equally concerning was what happened in the inference marketplaces. Multiple resellers could view and retain prompts using standard logging, without breaking any rules or exploiting vulnerabilities. The architecture of these routers means that every hop can see the full prompt and response, and the lowest-cost providers—sometimes operating on consumer plans or even stolen API keys—can scoop up the data as it passes through.




