Dragos: canaries catch data theft at three AI inference providers
Dragos seeded two million AI prompts with canary credentials and found three AI hosting providers leaking them to attackers.
· Dragos · Research by HD Moore
Dragos planted canary credentials in about two million AI prompts over five months and found attackers stealing data at three AI hosting providers, none of which knew about the theft until Dragos told them.
01What happened
HD Moore and the Dragos research team sent roughly two million prompts carrying unique canary URLs and credentials to AI services between July and November 2026. Three hosting providers turned out to be leaking that data to attackers, with the earliest incident dating to mid-July. Dragos says none of the three knew about the breach until it notified them.
02How it works
Each prompt carried a hostname, URL or credential that no legitimate system would ever use, so any contact with a Dragos listener proved the prompt had left its intended path. Dragos ran DNS, web, git, SSH, database and mail listeners, plus real AWS keys tied to empty, fully logged accounts. A second technique, which Dragos describes as a traceroute for AI inference, used an image URL and malformed requests whose nested error responses expose each hop, tracing about 2,000 routes.
03Who is affected
Customers of AI model routers, reseller marketplaces and smaller hosting providers are the exposed group, and Dragos names none of the providers. Prompts sent to United States providers that resell a Chinese lab's models still reached that lab, including requests marked zero data retention.
04What to do
Go directly to a first-tier provider or self-host where the budget allows, and verify that a router's privacy settings actually change where traffic goes. Seed your own canary credentials in agent and pipeline context and treat any use of them as an incident. Enable secret filtering where a provider offers it.
Questions people ask
- Which AI providers did Dragos find breached?
- Dragos does not name the three breached hosting providers in its write-up. The report says the earliest incident dates to mid-July 2026, that none of the three knew about the breach until Dragos notified them, and that one published a preliminary security notice after the late-October thefts. Customers of model routers and reseller marketplaces are the exposed group.
- What did attackers do with the stolen AWS keys?
- Attackers tested the Dragos bait keys within hours of taking them, then listed S3 buckets and searched Secrets Manager and Parameter Store. Dragos also saw attempts to enable Bedrock models, create Bedrock API keys and add IAM users. Across the study the bait keys were used for tens of thousands of AWS API calls, and some remained in use months later.
- Does a zero data retention setting stop prompts from being stored?
- Zero data retention did not reliably limit where prompts travelled in the Dragos study. A Chinese AI lab's systems fetched the secret URLs from more than half the prompts sent to its models between July and October 2026, including about two-thirds of requests marked zero data retention. The scraper typically appeared two to three days after the prompt.
- How can defenders tell whether their own prompts are leaking?
- Seeding canary credentials is the method Dragos recommends: place unique URLs and fake secrets in agent and pipeline context, then treat any contact with them as an incident. Dragos started with Thinkst Canarytokens before building its own service. HD Moore also advises turning on secret filtering where a provider offers it, while noting that filtering does not catch everything.
Sources 2 links, all fetched
- 1dragos.comDragos Research
- 2dispatch.thorcollective.comRelated Canary Study