# AI router test finds stolen keys and read prompts, even under ZDR

> Dragos planted two million tripwires in prompts; OpenRouter has since told customers their MiniMax requests may have been read.

URL: https://claudegpt.si/news/ai-router-canary-credentials-leak  
Published: 2026-11-21T22:56:23.257Z  
Updated: 2026-11-22T02:54:46.989Z  
Author: ClaudeGPT Si Newsroom  
Section: AI Tools  
Topics: OpenRouter, MiniMax, Dragos, AI security, Privacy  
Publisher: ClaudeGPT Si

## Key takeaways

- Dragos sent about two million prompts with unique fake credentials and links through AI services, routers and marketplaces from July to November 2026.
- It found three breached hosting providers and a Chinese AI lab that read links from about two-thirds of prompts marked for zero data retention.
- OpenRouter told customers their MiniMax requests may have been read and removed the ZDR label from those endpoints, Ars Technica reported.

A five-month test by the security firm Dragos found that prompts sent through AI routers and inference marketplaces were stolen, logged and read, including prompts marked for zero data retention. HD Moore, Dragos's chief research officer, [published the results on November 19](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens). On November 21, OpenRouter, the biggest AI router, began emailing customers that requests they sent to MiniMax models may have been read, [Ars Technica reported](https://arstechnica.com/security/2026/11/a-leaky-ai-supply-chain-is-exposing-secret-user-data-at-an-alarming-rate/).

An AI router takes a developer's request and passes it to whichever company hosts the chosen model, often the cheapest one. That means a prompt can travel through several companies the developer never signed a contract with. The test shows what can happen to it on the way.

## Two million prompts, each with a tripwire

From July to November 2026, Dragos sent about two million prompts to AI services, routers and inference marketplaces. Each prompt carried unique "canaries": fake credentials and secret web links that exist nowhere else. If anyone used one, Dragos knew exactly which request it came from. The AWS keys led to empty accounts with full logging.

Dragos traced about 2,000 routes. Several hundred passed through systems it couldn't identify, and some requests reached companies that weren't on the advertised path. One marketplace seller added roughly 1,000 hidden tokens to each prompt.

## Three breached providers, and ZDR prompts that were read

The findings fall into three groups.

- **Breaches.** Attackers stole data from three hosting providers, the earliest in mid-July. None of them knew until Dragos told them. In the largest, every request sent to one provider during late-October windows was stolen, including a couple hundred AWS keys. Attackers tested the keys within hours and made tens of thousands of AWS API calls, trying to turn on Bedrock models and add users. In the other two breaches, some stolen requests were marked zero data retention (ZDR), a setting meant to stop a provider from storing prompts at all.
- **Resellers.** Some sellers on an inference marketplace listed models at low prices, turned on request logging and collected the prompts. One router was quietly forwarding part of its traffic to that marketplace. Dragos says it stopped after being told.
- **Retention.** A Chinese AI lab fetched the secret links from more than half of the prompts sent to its models from July to October, including about two-thirds of requests marked "no data collection" or ZDR. U.S. hosts that resell the lab's models set off the same scraper, so prompts sent to them reached the lab too. "Zero Data Retention is supposed to mean something," Moore writes.

Dragos doesn't name the providers. It says it began notifying them in early November and is working with about a dozen.

## OpenRouter pulls the ZDR label from MiniMax models

OpenRouter's email names one of them. According to Ars Technica, which cited customer posts on Hacker News and X, OpenRouter said the researcher's planted links were opened whether prompts went to MiniMax, a Chinese model maker, directly or through three other hosts: AtlasCloud, Novita and StreamLake. OpenRouter has removed the "Zero Data Retention" label from the MiniMax-model endpoints while it investigates. The notice was not on OpenRouter's blog as of November 21, and MiniMax has not commented publicly.

This matters more than one vendor's slip. A ZDR label is what many companies rely on when they send code, contracts or customer data through a router. The test shows the label describes a policy, not what every host along the route does.

## What Dragos tells developers to do

Moore's advice, from the report:

- Self-host models if you can afford the hardware.
- Otherwise, use a first-tier provider with a single contract and privacy policy.
- With a router, pick the upstream providers yourself, check that privacy settings change the route, and avoid reseller marketplaces.
- Plant your own canary credentials in agent and pipeline contexts, and treat any use as an incident.

Anyone who sent secrets through MiniMax models on OpenRouter since June should treat them as exposed and rotate them. Dragos says canaries planted in July were still firing when it published, so the leaked data is still being used.

Update, November 22: OpenRouter has now [posted its notice](https://openrouter.ai/blog/announcements/data-processing-incidents-on-inference-providers/). It off-boarded three providers it did not name, removed 125 endpoints from zero data retention eligibility and notified potentially affected customers. Novita said some MiniMax requests sent between May 28 and November 11 went to MiniMax's own API, Phala said an attacker controlled the server running DeepSeek V4 Flash 0731 from October 21 to November 8, and Together AI said prompt samples from June and July were downloaded after an AI coding agent committed a credential to a public repository. Our [full report](https://claudegpt.si/news/openrouter-offboards-providers-zdr) has the details.

## Sources

- [Three breaches, endless leaks: Your AI supply chain needs canaries](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens), Dragos
- [Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.](https://arstechnica.com/security/2026/11/a-leaky-ai-supply-chain-is-exposing-secret-user-data-at-an-alarming-rate/), Ars Technica
- [Data Processing Incidents on Inference Providers](https://openrouter.ai/blog/announcements/data-processing-incidents-on-inference-providers/), OpenRouter
