# Researcher seeds 2 million canary secrets into AI routers and finds data leaking

> HD Moore planted 2 million traceable fake secrets in prompts sent through AI routers and reports that sensitive information leaked widely along the chain.

- Status: reported (One outlet, not independently verified yet.)
- Topic: Safety & Security
- Published: 2026-11-20T20:59:08+00:00
- Updated: 2026-11-21T07:54:39.663689+00:00
- Web page: https://ai-news.ikangai.com/story/researcher-seeds-2-million-canary-secrets-into-ai-routers-and-finds-data-leaking/
- Seen in: Ars Technica
- Change: security (Security)
- Applies to: OpenRouter, Vercel, AWS Bedrock

## What changed

Researcher found prompts sent through AI routers are decrypted at each hop, and seeded secrets were stolen and misused by attackers at three hosting providers and by a Chinese AI lab's systems.

HD Moore seeded roughly two million fake but working secrets into prompts sent through AI routers. He found that the data was taken and misused at several points along the chain, including by attackers who broke into hosting providers. The test used AWS access keys tied to empty accounts, SSH passwords, unique URLs, DNS names and private git repository references. None of them had been used before, and Moore controlled the systems that received any request using them. Each secret went into exactly one prompt sent to a single provider. Any later use of it could be traced back to the link that exposed it. That is the logic of a canary: a value that exists in only one place, so any appearance of it marks that place as the source of the leak.

AI routers, also called AI gateways, sit between an app and the models that answer its requests. Services such as OpenRouter and Vercel accept a prompt through one API endpoint and forward it to whichever inference provider fits a criterion, such as lowest price or suitability for the task, with unified billing. As the Dragos write-up explains, Moore's point is that prompts are encrypted in transit but decrypted at each hop, because the router must read the plaintext to choose a destination. A prompt can also pass from one router to another, and each party may log it for as long as its policy allows. Moore wrote that every hop can see the entire prompt and response.

The most serious finding was a breach of three hosting providers. Moore wrote that attackers were actively stealing data from them, and that none of the three providers knew about the breach until he notified them. The largest incident involved a couple hundred of his AWS keys, each sent in one request to one provider within a few short windows in late October. Some keys came from requests the provider rejected before any model ran. Moore said this shows the attacker was copying requests at the provider's front door rather than deeper in the inference stack. In the other two breaches, some stolen requests were marked Zero Data Retention. Moore wrote that a privacy flag does not help when the attacker is already inside. Once they had the keys, the attackers tested them, searched S3 buckets and secrets stores, tried enabling hosted models in AWS Bedrock, and added their own IAM users. Moore wrote that the bait keys were used for tens of thousands of AWS API calls. One adversary appears to have compromised two providers at once, since secrets sent to both were used together from a single rented server.

A second finding needed no exploit. Moore wrote that some marketplace sellers listed a model at the lowest price, turned on request logging, and waited. He identified them by sending the same model's requests through different sellers. When a credential was reused only after passing through one seller, that seller was the likely source, not the model provider. Moore described these people as individuals working from what looked like home connections. One of them tried to log in over SSH by hand, using a password from a prompt.

The third finding involved companies fetching secrets they were never meant to see. Moore wrote that a Chinese AI lab's systems fetched the secret URLs from more than half of the prompts sent to its models between July and October. That included roughly two-thirds of requests marked as no data collection or Zero Data Retention, typically two or three days after the request. Several US hosting providers resell that lab's models. Prompts sent to them triggered the same scraper, which Moore said meant the prompts were flowing back to the lab. Moore also described one incident in which a seeded link was posted into a Feishu chat and opened from devices in Beijing and Shanghai more than three weeks after the prompt went to a US-based provider with the ZDR flag. He said the disclosure to that provider included only the first eight characters of the link, which he took to mean the sharer was reading a stored copy of the full prompt.

Moore's advice, as Ars Technica reported it, is that the most effective protection is self-hosted models on on-premises servers, which he acknowledged is expensive. For others, he suggested a first-tier provider with one contract and a privacy policy. He noted that even then nothing guarantees no leak. Users should avoid default provider or reseller pools and choose upstream providers themselves. Dan Goodin, Ars Technica's senior security editor, described the privacy and security practices across much of the AI ecosystem as lax. He said people who want their secrets kept private face a daunting task that isn't likely to become easier soon.

## Why it matters

Companies that send prompts through third-party routers or gateways may expose credentials and private data to intermediaries they do not directly control. The experiment suggests such leaks can occur along the chain of services.

## What you can do

Check whether your AI apps send secrets or sensitive data through routers or gateways, and keep credentials out of prompts.

## Original source

- [AI Supply Chain Security: What Canary Tokens Found | Dragos - Dragos](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens)

## Further reporting

- [Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced. - Ars Technica](https://arstechnica.com/security/2026/11/a-leaky-ai-supply-chain-is-exposing-secret-user-data-at-an-alarming-rate/)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Researcher seeds 2 million canary secrets into AI routers and finds data leaking", 2026-11-20, https://ai-news.ikangai.com/story/researcher-seeds-2-million-canary-secrets-into-ai-routers-and-finds-data-leaking/
