Canary test finds AI routers and hosts leaking prompts and API keys
A researcher seeded nearly two million prompts with unique trackable secrets and found breaches at three hosting providers, resellers logging traffic, and providers ignoring zero-retention flags.
TL;DR
- Researcher HD Moore planted nearly two million traceable fake secrets in AI prompts sent through routers and providers.
- Attackers stole prompts at three hosting providers, and some stolen requests were marked zero data retention.
- Developers who route AI traffic through brokers should check who actually handles their prompts and secrets.
How the test worked
Many AI apps do not call a model vendor directly. They send requests to AI routers, also called AI gateways, such as OpenRouter or Vercel. A router reads the prompt in plain text, picks an inference provider based on factors like price or task fit, and forwards the request. Prompts are encrypted in transit, but they are decrypted at each hop, so every party in that chain can see them.
To measure what happens to that data, Moore, chief research officer at security firm Dragos, built a set of "canaries". These were AWS access keys tied to empty accounts, SSH passwords, and unique URLs, DNS names and private git repository references. None had been used before, and he controlled the servers that would receive any use of them. Each secret went into exactly one prompt sent to one provider, so any leak could be traced to a specific link in the chain.
Hosting providers were breached
The most serious finding was that three separate hosting providers were compromised. In the case Ars Technica described in most detail, attackers reused a couple hundred of the AWS keys. Some keys were copied from requests the provider rejected before any model ran, which suggests the attacker was capturing traffic at the provider's entry point rather than deeper in the inference stack. In the other two breaches, some of the stolen requests had been marked Zero Data Retention (ZDR), a flag meant to stop providers from keeping data. As Ars Technica put it, the flag "doesn't help when the attacker is already inside."
The stolen keys were used "tens of thousands" of times in AWS API calls. The attackers searched S3 buckets, secrets managers and parameter stores for more credentials, tried to enable hosted models in AWS Bedrock, and added their own IAM users to keep access. Keys sent to two of the providers were used almost immediately from the same rented server, which points to a single adversary behind both breaches.
Resellers logged prompts
The test also found resellers in the inference marketplace that logged prompts and passed secrets on to others. Some offered low prices to attract traffic, then turned on request logging. Moore could tell them apart by sending the same model requests through different sellers and checking which seller's credentials were reused later.
Providers fetched secret URLs
The third problem was providers ignoring privacy settings. The clearest case involved a Chinese AI lab whose systems fetched the secret URLs from more than half of the prompts sent to its models between July and October, including about two-thirds of requests marked "no data collection" or ZDR. The fetches typically came two or three days later. Several US hosting providers resell this lab's models, and prompts sent to them triggered the same scraper. Moore also reported that one US host crawled a secret URL more than two months after a single call, and that a git client on a different cloud tried to clone private repositories referenced only in prompts sent to this lab's models.
Why it matters
For developers, the practical lesson is that a privacy flag is only as strong as the chain behind it. Each router, reseller and host may log and store a prompt for as long as its own policy allows. Moore said the most effective protection is self-hosting models on your own servers, which is costly. The next best option, he said, is a first-tier provider offering "one contract and a privacy policy to read", which reduces risk but does not remove it. He also advised against accepting default provider or reseller pools and choosing upstream providers directly.
Any secret that appears in a prompt, such as a cloud key, password or private repository URL, should be treated as potentially exposed. Rotate credentials that have been sent to AI services, keep them out of prompts, and use scoped, short-lived keys wherever possible.
Sources
This article was written by Blozhik from the reporting below. Read the originals for full detail.
ArResearcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.Ars Technica