AI Canary Study Finds Evidence of Three Provider Breaches and Prompt Theft
Emerging Technology Risk · 7 sources · Updated 10h ago · First seen Nov 19, 2026

Dragos researcher HD Moore reported that approximately two million AI prompts seeded with unique tracking URLs and canary credentials uncovered evidence of compromises at three unnamed AI hosting providers, unauthorized access to customer prompts by marketplace resellers, and apparent violations of privacy settings. The study, conducted between July and November 2026, tracked attackers using planted AWS credentials in tens of thousands of API calls to inspect cloud resources, attempt to enable additional services, and establish persistent access. The largest reported compromise captured requests during several short windows, including requests rejected before model execution. Resellers also allegedly accessed prompts through built-in logging. Dragos
The findings highlight risks from opaque AI routing chains: prompts submitted through U.S. providers reportedly reached an unnamed Chinese AI lab, which fetched secret URLs even from requests marked Zero Data Retention or “no data collection.” These observations raise concerns about confidentiality and data-residency assurances, although canary activations do not necessarily establish which intermediary leaked a prompt, and the unnamed providers limit independent verification. Dragos said affected providers were notified. Recommended mitigations include filtering secrets from prompts, minimizing intermediaries, avoiding reseller marketplaces, explicitly selecting upstream providers, and monitoring seeded canaries; on-premises inference is another option where feasible. Ars Technica


