Researcher plants 2m canaries in AI supply chain
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Security researcher HD Moore seeded nearly two million trackable canary credentials into AI prompts from July to November 2026.
The study points at the brokers between users and models, not the model makers themselves.
Background
HD Moore is a security researcher at Dragos, Inc., the company whose team ran the experiment. Many AI apps don't talk to a model directly; they hand prompts to brokers such as OpenRouter or Vercel, which collect large volumes of prompts and forward each to whichever inference engine fits criteria such as lowest price. Moore wanted to know what happens to a secret once you type it into an AI app, so he planted roughly 2 million digital tripwires in prompts and sent them through that network. The industry has spent energy reassuring customers about what big model makers do with their data; Moore's work points to the brokers, resellers and hosting providers in between.
How it unfolded
- Jul 2026Earliest confirmed theft activity dates to mid-July 2026, meaning leaks were underway almost as soon as the experiment began.
- Jul 2026 to Nov 2026Moore ran the study, embedding nearly two million unique, trackable canary credentials and URLs into requests sent to AI inference services and gateway routers.
- Late Oct 2026Approximately 200 unique AWS credentials were exfiltrated during targeted attacks.
- Nov 2026Affected providers learned about the breaches only when Moore notified them.
- Nov 19, 2026Moore's report was publicly summarized.
- Nov 20, 2026Stories on the experiment were published by Ars Technica, CryptoBriefing, HyperAI and others.
Key terms
- canary credentials
- Unique fake login details planted so that any leak or misuse can be traced back to its source.
- AI supply chain
- A US-led international initiative to secure supply chains for semiconductors, AI, and rare earth elements.
- inference engine
- The process by which an AI system applies rules to known facts to derive new conclusions.
Source: Wikipedia
Who’s saying what
- Expert
- Moore said he wanted to know what happens to a secret once you type it into an AI app, and planted the tripwires to find out.
- Caution
- Chris H. of ResilientCyber said the findings should worry anyone sending sensitive data to AI services, and that a privacy flag doesn't help when the attacker is already inside.
Still unverified
HyperAI's report states that no company, location or date was disclosed in the available report.