---
title: Dragos seeded two million AI prompts with canary secrets and found three breached AI hosting providers, resellers logging prompts and privacy flags ignored | ctxsec
url: https://ctxsec.com/items/dragos-canary-study-three-ai-hosting-providers-breached-resellers-logging-prompts-zdr-ignored
description: "From July to November 2026, Dragos sent about two million prompts containing unique URLs and canary AWS credentials to widely used AI services and watched where the secrets showed up. It says attackers were stealing requests from three unnamed AI hosting providers, the earliest from mid-July, and none knew until Dragos told them. In the largest case a couple hundred bait AWS keys were taken, tested within hours and used for tens of thousands of AWS API calls, including attempts to enable Bedrock models and add IAM users. Dragos also found inference-marketplace resellers viewing and keeping customer prompts through ordinary request logging, and providers that appeared not to honor zero-data-retention or privacy settings, including a secret link from a ZDR request that was posted in a Feishu chat 24 days later."
status: "research (Dragos blog by HD Moore, Nov 19, 2026); providers not named; findings are Dragos's own monitoring and not independently verified"
published: 2026-11-19
---

> From July to November 2026, Dragos sent about two million prompts containing unique URLs and canary AWS credentials to widely used AI services and watched where the secrets showed up. It says attackers were stealing requests from three unnamed AI hosting providers, the earliest from mid-July, and none knew until Dragos told them. In the largest case a couple hundred bait AWS keys were taken, tested within hours and used for tens of thousands of AWS API calls, including attempts to enable Bedrock models and add IAM users. Dragos also found inference-marketplace resellers viewing and keeping customer prompts through ordinary request logging, and providers that appeared not to honor zero-data-retention or privacy settings, including a secret link from a ZDR request that was posted in a Feishu chat 24 days later.

[← AI](/ai) · [markdown](/items/dragos-canary-study-three-ai-hosting-providers-breached-resellers-logging-prompts-zdr-ignored.md)

research (Dragos blog by HD Moore, Nov 19, 2026); providers not named; findings are Dragos's own monitoring and not independently verified

2026-11-19 7:35 PM ET

### Dragos seeded two million AI prompts with canary secrets and found three breached AI hosting providers, resellers logging prompts and privacy flags ignored

- **Never put live secrets in prompts** or in files agents send to models; assume anything sent to an inference provider or router can be logged or stolen.
- **Pin providers** for sensitive work instead of letting a router pick the cheapest seller, and check each provider's data-retention terms.
- **Seed your own canaries:** add canary tokens to test prompts and RAG data so you find out if a provider leaks them.
- **Watch AWS for** Bedrock model-access changes, new Bedrock API keys and new IAM users you didn't create.

Sources: [Dragos (HD Moore): Three breaches, endless leaks: Your AI supply chain needs canaries (Nov 19, 2026)](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens)
