A five-month security study using nearly two million instrumented AI prompts found evidence of breaches at three hosting providers, according to research published by industrial cybersecurity company Dragos. The researchers also reported prompt theft and cases in which requests marked for Zero Data Retention were exposed.

From July through November 2026, the team placed unique URLs and synthetic cloud credentials inside prompts submitted to widely used AI services. These canary tokens were not meant for legitimate use. Any later attempt to visit a URL or use a credential would indicate that someone beyond the expected inference process had obtained the prompt.

Dragos says the most serious signals came from three providers whose systems appeared to be compromised. In the largest incident, hundreds of planted AWS credentials were accessed after being sent during several windows in late October. The researchers observed that some requests had been rejected before a model ran, leading them to conclude that copying occurred near the provider’s entry point rather than deeper in model execution. They said none of the three providers knew about the intrusions before notification.

The bait credentials generated tens of thousands of AWS API calls. According to the report, actors tested whether keys worked, searched services including S3, Secrets Manager and Parameter Store, attempted to enable hosted models, and tried to create additional access for persistence. In two incidents, stolen prompts had carried ZDR flags, illustrating that a retention policy cannot prevent theft when an attacker already has access to the processing environment.

The study focused on a supply chain that can involve routers, marketplaces, model developers, cloud companies and independent inference hosts. Every intermediary that can read a prompt becomes part of the security boundary. Price-driven routing can make that boundary harder to see because customers may interact with one endpoint while their data is forwarded through other services before reaching a model.

Dragos’s results are based on monitored canaries rather than a complete measurement of all customer traffic. Still, the controlled design gave each planted secret a known origin and route, allowing the researchers to link later activity to a particular request. The company said real customer prompts were processed alongside its bait and were likely at risk where theft occurred, though the supplied evidence does not quantify any confirmed customer losses.

For organizations using third-party AI inference, the research supports treating prompts like other sensitive outbound data. Providers’ retention settings remain useful, but they do not replace vendor assessment, route visibility, credential hygiene and monitoring for unexpected access. Canary tokens can add an independent signal when policy statements alone cannot show what happens after a request leaves the customer’s environment.