Latest
AI Affairs, home

Saturday 21 November 2026

News

AI prompt test finds three compromised hosts and leaking resellers

Nearly 2 million prompts carrying trackable secrets exposed stolen customer data, reseller logging and URL fetching that continued despite privacy settings.

Aerial view of AWS datacenters in Umatilla, with another building under construction
Photo: Tedder, CC BY-SA 4.0, via Wikimedia Commons (cropped)

HD Moore, chief research officer at security firm Dragos, disclosed on 19 November that attackers had stolen prompts sent to AI hosting providers, Ars Technica reported. His experiment placed a unique, trackable secret in each of nearly 2 million prompts sent through AI services and found evidence of breaches at at least three hosting providers. The test covered routers such as OpenRouter and Vercel, which direct requests from apps towards inference engines, as well as resellers and hosts.

Key points

  • Moore found evidence that attackers siphoned prompts from at least three AI hosting providers.
  • Stolen test credentials were used tens of thousands of times in AWS API calls.
  • Some resellers logged prompts despite zero-data-retention settings.
  • A Chinese AI lab’s systems fetched secret URLs from prompts, including requests marked for no data collection.

Stolen AWS keys at three hosts

Moore used previously unused AWS access keys, credentials tied to empty AWS accounts, passwords and private URLs as canaries. Each appeared in just one prompt sent to one provider, allowing him to trace subsequent attempts to use it.

At one compromised host, attackers took every request Moore sent during several short periods in late October and later used a couple hundred of the planted AWS keys. Some stolen requests had been rejected before a model ran, Moore said. In the other two breaches, some affected requests carried zero-data-retention settings.

The stolen keys generated tens of thousands of AWS API calls. The parties using them searched S3 buckets, secrets managers and parameter stores, tried to enable models in AWS Bedrock and added IAM users to retain access.

Reseller logging despite privacy settings

Moore found that multiple inference resellers logged prompts and passed them to people who used the embedded credentials. He compared requests to the same model through different sellers: credentials sent through one seller were later used, while those sent through another were not.

Chinese lab fetched secret URLs

Between July and October, a Chinese AI lab’s systems fetched secret URLs from more than half the prompts Moore sent to its models, including roughly two-thirds of requests marked for no data collection or zero data retention, Moore wrote. Requests sent to US hosting providers reselling the lab’s models triggered the same scraper.

The affected hosting providers learned of the breaches when Moore notified them in November 2026, Crypto Briefing reported.

Sources

Topics: Inference, Safety

What should we cover? Tell us

The weekly briefing

The AI Affairs weekly briefing: the week's 10 most important AI stories in business, government and markets, every Monday. We'll email you a link to confirm. Every email has a one-click unsubscribe. Privacy policy