Dragos canaries show prompt theft at three AI hosts
Original title: Three breaches, endless leaks: Your AI supply chain needs canaries
About two million seeded prompts from July through November 2026 exposed breaches, reseller logging and ignored privacy settings.
Dragos researchers, led by HD Moore, sent nearly two million prompts with unique URLs and canary credentials to widely used AI services and watched where they surfaced. They report attackers stole data from three hosting providers that did not know until notified, including a late-October case in which a couple hundred AWS keys were taken at one provider's front door. Separate inference-marketplace resellers could view and keep prompts through ordinary logs, and some Zero Data Retention requests were still stolen or forwarded, including to Chinese organizations.
Unconfirmed: The three hosting providers are not named, and copying of real customer prompts is inferred from bait processed alongside them.
- The largest case copied every request in short late-October windows, including rejected ones.
- Two of the three breaches included requests marked Zero Data Retention.
- One seeded link later appeared in an internal company chat in China.
- Unique URLs were accessed months after a single inference request.
Impact: Customer prompts and seeded cloud credentials copied from AI hosts
- ATT&CK
- T1078 Valid Accounts
- Organizations
- Dragos
- Countries
- ChinaUnited States
From the original article
13 min read at dragos.com
HD Moore, Chief Research Officer, will be regularly sharing work from the research team he leads at Dragos. Topics will include xOT, IT, LLMs, PQC, and more. Their findings will sharpen defenders’ understanding of exposures and vulnerabilities emerging across the threat landscape.
By dragos.com via wonk · dragos.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.