On November 19, 2026, Dragos chief research officer HD Moore published the results of a five-month canary study. From July through November the team submitted approximately two million prompts, each carrying unique URLs and credentials, to widely used AI services and watched for anything that touched them. CYBR.SEC.Media summarized the report on November 20. No provider is named in either source.
Three distinct failure modes
Provider compromise. Dragos reports evidence that data from at least three hosting providers was stolen, with the earliest incident in mid-July, and says none of the three knew before being notified. In the largest case a couple hundred bait AWS keys, each sent once, were taken during a few late-October windows, including from requests rejected before a model ran. Dragos reads that as copying at the provider's front door. Attackers tested keys, listed S3 buckets, searched Secrets Manager and Parameter Store, and tried to enable Bedrock models and add IAM users, adding up to tens of thousands of API calls. One actor later used an LLM-driven pentest framework whose user agents Dragos publishes for detection.
Marketplace resellers. Sellers on an inference marketplace could read and keep prompts through ordinary request logging, with no exploit needed.
Privacy settings not honored. Dragos says one Chinese lab's systems fetched secret URLs from more than half of the prompts sent to its models between July and October, including roughly two-thirds of requests marked no-data-collection or Zero Data Retention, and that US resellers of that lab's models triggered the same scraper. A link sent with a ZDR flag appeared in a company chat 24 days later.
What the study does not prove
Dragos observed only its own bait. Its statement that real customer prompts were almost certainly copied as well is an inference from shared processing, not a measured result. The report says it is working with about a dozen providers and that one has published a preliminary notice.
Treat the prompt path as part of your system
Analysis: a ZDR flag is a request to whoever honors it, not a control you enforce. Each router hop sees the whole prompt. Credentials should never appear in prompts, and outbound context should be filtered as strictly as tool output.
Original exercise, not executed: in a staging copy of the Belgavi's Book Store help desk, Tanishka adds one unique, unused URL per routed request for CASE-5120 test traffic, logging the router, model and privacy flag alongside it. Any later hit on a URL identifies the path that leaked it. Use real empty accounts with full logging, never live keys.
Sources accessed November 22, 2026. Views are my own and do not represent any employer, past or present.
Sources and evidence
Dragos's canary study reports three breached inference providers, prompt-reading resellers and ignored ZDR flags. Keep secrets out of prompts, know every hop on your routing path, and treat privacy flags as requests rather than controls.