# Dragos plants 2M canaries in AI prompts — three providers found breached

> Two million booby-trapped prompts show where AI requests really go, and who reads them on the way.

- Org: Dragos
- Date: 2026-11-19
- Category: security, article
- Source: https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens
- Canonical page: https://ai-tldr.dev/releases/dragos-ai-supply-chain-canaries/

## In short

Dragos researcher HD Moore sent about two million prompts with unique fake secrets to AI services. Attackers stole keys from three breached providers, resellers logged prompts, and some hosts ignored Zero Data Retention flags.

## What it is

Dragos published a five-month canary experiment on the AI inference supply chain. HD Moore seeded unique AWS keys, passwords, DNS names, URLs and private git repo references into about two million prompts, each secret sent exactly once to one provider, and then watched which ones were later used.

## How it works

Every canary is tied to the exact prompt, model, router and privacy setting that carried it, so a reused secret points at the hop that leaked it. Dragos also traced about 2,000 inference routes by embedding call-back image URLs and sending slightly malformed requests that expose each hop's error envelope. Several hundred routes included a system that was not on the advertised path.

## Why it matters

The study found attackers inside three hosting providers, marketplace resellers that logged prompts and reused the secrets, and a Chinese AI lab whose systems fetched secret URLs from more than half of the prompts sent to its models, including about two-thirds of requests marked no data collection or ZDR. Anything an agent puts in a prompt, including credentials, can travel far beyond the provider you picked.

## Who it's for

developers using AI routers, security teams, anyone sending secrets through agents

## Try it

Put a free Thinkst Canarytoken into the context your agents send and treat any use of it as an incident.

## Quick facts

- Researcher: HD Moore, Chief Research Officer, Dragos
- Test period: July to November 2026
- Prompts seeded: About 2 million, each with unique canary secrets
- Providers breached: At least 3, none aware until Dragos told them
- Bait-key abuse: Tens of thousands of AWS API calls
- Follow-up: OpenRouter emailed customers about MiniMax-model prompts

## FAQ

### Which AI providers were named in the Dragos canary study?

The Dragos post anonymizes most providers. It names OpenRouter, found through route tracing, and a marketplace seller called SurplusIntelligence. After the research, OpenRouter emailed customers that prompts sent to MiniMax models, directly or through AtlasCloud, Novita and StreamLake, may have been read, according to Ars Technica and PrivacyAffairs.

### Does Zero Data Retention protect my prompts on AI routers?

Not reliably, according to Dragos. In two of the three provider breaches some stolen requests were marked ZDR, and a Chinese lab's scraper fetched URLs from ZDR-flagged requests. HD Moore also notes that privacy flags apply only to the first hop, and OpenRouter removed the ZDR label from MiniMax-model endpoints while it investigates.

### What should I do if my prompts went through an affected AI provider?

PrivacyAffairs advises treating affected prompts as read and revoking, not just replacing, any password, API key or token that appeared in them. That includes signed download links, webhook URLs, password-reset links and internal hostnames. Check your OpenRouter activity log even without the notice email, because the reported time windows do not all match.

### How can teams reduce AI supply chain leaks according to Dragos?

Dragos lists five steps: self-host models for the most sensitive work, call a first-tier provider directly under one contract, choose router upstreams yourself and avoid reseller marketplaces, seed canary credentials into agent context, and turn on provider-side secret filtering. HD Moore calls secret filtering helpful but not enough, since much sensitive data has no key-like format.

## Links

- [Dragos Research](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens)
- [Ars Technica](https://arstechnica.com/security/2026/11/a-leaky-ai-supply-chain-is-exposing-secret-user-data-at-an-alarming-rate/)
- [OpenRouter Notice](https://www.privacyaffairs.com/openrouter-minimax-prompts/)
- [HN Discussion](https://news.ycombinator.com/item?id=50585962)
- [Techmeme](https://www.techmeme.com/261121/p10#a261121p10)
