egov.mn
Монголоор унших

Two million planted secrets reveal AI supply chain leaking user prompts, Dragos researcher finds

Dragos researcher HD Moore sent nearly two million AI prompts seeded with unique secrets through popular AI services and found three compromised providers, resellers logging customer prompts and widespread disregard of zero-data-retention settings.

This story was prepared automatically by software.

Illustration of a data cube set as bait in a trap, representing AI privacy leakage
Illustration of a data cube set as bait in a trap, representing AI privacy leakage

Sensitive information sent to AI models can leak extensively as it passes through the chain of routers, resellers and hosting providers behind many AI apps, according to research by HD Moore, chief research officer at security firm Dragos. Moore published the findings on November 19, and Ars Technica reported on them in detail.

How the experiment worked

From July through November 2026, Moore's team sent nearly two million prompts containing never-before-used "canary" secrets, including working AWS access keys tied to empty accounts, unique URLs, DNS names and private git repository references, to widely used AI services and the "AI routers" that broker traffic among them. Each secret was seeded into exactly one prompt sent to a single provider, so any later use pointed to the link in the chain responsible.

Key findings

The research reported three main problems:

  • Provider breaches: Attackers had compromised at least three hosting providers and were copying incoming requests. In the largest case, a couple hundred AWS keys were stolen, some tested within hours, and the bait keys were used in tens of thousands of AWS API calls. None of the providers knew about the breaches until notified.
  • Resellers reading prompts: Some sellers on an AI inference marketplace listed models at the lowest price, switched on standard request logging and retained customer prompts, without exploiting any vulnerability.
  • Privacy settings ignored: A Chinese AI lab's systems fetched secret URLs from more than half of the prompts sent to its models, including roughly two-thirds of those marked "no data collection" or Zero Data Retention. Prompts sent through US hosting providers reselling that lab's models triggered the same scraper.

OpenRouter notifies customers

In a November 21 update, Ars Technica reported that AI router OpenRouter emailed customers that some of their prompts may have been read, citing the research. OpenRouter said the exposure involved prompts sent to China-based model provider MiniMax, directly or through three other resellers, and removed the Zero Data Retention label from MiniMax endpoints.

Moore said the most effective protection is to host models on premises or, failing that, to work directly with a first-tier provider under a single contract and privacy policy, and to avoid placing secrets in prompts.

Sources: Dragos, Ars Technica

Share
Next storyAmazon closes in on $7.5 billion deal to buy AI startup Decart
Newsletter

Get the latest news delivered to your inbox.

Related articles