Fake credentials planted inside about two million AI prompts were used by outsiders, and the researchers who planted them tracked every request. The trail ran through AI routers, the middlemen developers use to reach model providers. The most visible fallout so far is an OpenRouter prompt theft case that ended with three providers dropped from the service.
Routers, also called AI gateways, accept a prompt from a customer and pass it to an upstream provider. The customer usually sees one endpoint and rarely the chain behind it. That gap sits at the center of a red-team style study that security researchers at Dragos published on November 19.
What the Dragos canary study found
Canary tokens work like a smoke detector. A team plants a fake credential or a unique web link where only a monitored system should ever see it, then waits for someone to use it. HD Moore, who ran the experiment described in this Cryptonews report, seeded about two million prompts with unique links and credentials between July and November 2026.
He controlled the infrastructure that received any request made with those secrets. He also placed each secret in exactly one prompt sent to a single provider, so any later use pointed back to one path. Dragos published its findings as AI Supply Chain Security: What Canary Tokens Found, and disclosures continued through November 21 and 22.
The timeline below shows how the pieces fit together.
| Date | What happened |
|---|---|
| July to November 2026 | About two million canary prompts seeded with links and credentials |
| July to October 2026 | Test URLs reportedly placed in prompts sent to MiniMax models, according to a posted OpenRouter email |
| Late October 2026 | Approximately 200 AWS credentials reportedly exfiltrated in targeted attacks |
| November 19, 2026 | Dragos publishes the canary study |
| November 21, 2026 | A user posts what they describe as an OpenRouter email to Hacker News |
| November 22, 2026 | Researchers say planted keys and secret links are still being accessed |
Which providers were breached in the OpenRouter prompt theft
The study describes three breached providers and one marketplace seller. Only OpenRouter is named, and the rest appear only as anonymized descriptions. That leaves developers unable to check whether their own provider is on the list. The mechanics are still concrete enough to matter:
- One attacker appears to have compromised two of the providers at once, using secrets from both from a single rented server.
- One leaked key traveled through a different router that quietly forwarded some of its traffic to the marketplace. No customer was told about that hop.
- A Chinese AI lab’s 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, according to ai-tldr’s summary of the study.
The hidden hop deserves the most attention. Customers choose a router and read its privacy terms, but one forwarding step outside that agreement can move a prompt somewhere nobody agreed to send it. The Chinese lab result is likely to be read against a wider debate in Washington, where a House panel says US officials approved nearly 900,000 AI chips for China. Moore also traced prompts from US users back to Chinese AI organizations, according to one report of the study.
Why Zero Data Retention labels did not hold
Zero Data Retention, usually shortened to ZDR, is a setting that asks a provider not to store a prompt. Developers rely on it because it is easy to switch on and easy to explain to a compliance team. The study complicates that comfort. Some of the stolen requests were flagged ZDR, and the researchers wrote that “a privacy flag binds only the first hop.” If that holds, the setting protects the first stop on a route and says little about the stops after it.
OpenRouter’s reported response points the same way. According to an OpenRouter email that a user posted to Hacker News on November 21, and as summarized in a DEV Community post, the company removed the ZDR label from the affected MiniMax endpoints. The poster said the email appeared authentic, but no primary OpenRouter statement on the matter surfaced in the reporting reviewed here.
Enterprises face the sharpest exposure. Many compliance documents cite ZDR or “no data retention” labels, and the study says those labels did not reliably stop leakage. Legal and security teams should ask each provider which parties touch a request, not only whether a flag exists. Retention periods can also differ from what a toggle suggests. Google, for instance, now keeps Gemini chats for two months even with history turned off, according to a USA 24 report.
The disputed AWS key claims
The most alarming number in the study is about 200 AWS credentials. One report, published by Mezha, says attackers who compromised one provider reused “a couple hundred” AWS keys. A separate account says approximately 200 unique AWS credentials were exfiltrated during targeted attacks in late October 2026. Reports based on the Dragos work also describe stolen AWS keys used against S3 and IAM, the storage and access-control services that sit at the heart of many cloud accounts.
The problem is that a reported OpenRouter email says none of the AWS test keys were used. Those accounts are hard to reconcile. The email reached the public secondhand, through a Hacker News post, and no primary OpenRouter statement on the AWS question appeared in the reporting reviewed for this article. Until a primary source settles it, treat any AWS key that passed through an affected router as possibly exposed.
What developers and companies should do this week
Dragos closed its blog post with a short warning: “Use routers with care.” The steps below put that advice into practice.
- Check your OpenRouter activity log for July through November, even if you never received a notice. The reported time windows do not all match, so a clean log for one period does not clear the others.
- Choose upstream providers yourself instead of accepting the default pool. One write-up of the story advises staying away from reseller marketplaces, some of whose capacity runs on stolen keys.
- Rotate AWS and other cloud keys that passed through an AI router during the affected windows, then review S3 and IAM activity for anything you cannot explain. This recommendation follows from the leak findings. It is not a quoted advisory from Dragos, OpenRouter or a government agency.
- Keep secrets out of prompts. Developers often paste API keys into AI tools while debugging or automating work, so store credentials in a secrets manager and send the model only the code or data it needs.
- Treat ZDR as a contract term, not proof of safety. Ask providers which hops handle a request and whether any traffic is forwarded along the way.
Open questions the study has not answered
Several gaps remain, and each one changes what a careful developer should do.
- Which hosts were breached. Dragos anonymized most of the providers, so readers cannot yet tell which services were hit.
- Whether real customer data was taken. Dragos said real customer requests processed alongside the bait were “almost certainly copied too.” That is an inference from the study design, not a verified count.
- Whether access has stopped. Researchers say they are working with about a dozen providers, and planted keys and secret links were still being accessed months after the first thefts, as of this week.
- How regulators will respond. No US government advisory on the study turned up in the reporting reviewed here.
What comes next for AI routers
More names may follow. The researchers say they are working with about a dozen providers, which leaves room for further disclosures and could identify some of the anonymized hosts. The AWS dispute needs a primary statement from OpenRouter or the providers involved, rather than secondhand posts. Watch whether enterprise buyers start asking routers for documented data-flow maps, and whether the first-hop finding pushes router companies to publish their routing paths.
Lawmakers are already pressing on AI security. Senators asked the White House to act on AI agent cyber threats in a recent letter, and this study gives that debate a concrete example of supply-chain exposure (Senators Urge White House to Act on AI Agent Cyber Threats). Until the providers are named and the AWS question is settled, the safer assumption is that a prompt may travel farther than its settings say.




