Sunday, 22 November 2026

Is OpenRouter safe? What the 2 million canary test found

OpenRouter says its own systems were not breached, but prompts it passed to outside AI providers were copied, read and used, even ones marked never stored.

A server room aisle with one amber cable standing out among blue network cables

The short version

  • OpenRouter removed three providers and took Zero Data Retention status away from 125 endpoints on November 22.
  • The company says the research showed no breach of OpenRouter itself, only of providers it sends prompts to.
  • Researchers hid unique secrets in about two million prompts and watched who used them between July and November.
  • Phala, Novita and Together AI have each published notices describing separate ways prompts were exposed.
Quick read · 1 min

OpenRouter says its own systems were not breached, but prompts it forwarded to outside AI providers were exposed. On November 22 it removed three providers, took Zero Data Retention status away from 125 endpoints and said it had notified every customer who may be affected.

The proof came from the security firm Dragos, which hid unique secrets in about two million prompts and watched who used them. Stolen cloud keys were tested within hours in some cases, and prompts marked as never stored were kept and read by people.

  • Phala, Novita and Together AI have published notices about separate incidents.
  • Phala says affected users should change every key, token and password that appeared in their requests.
  • OpenRouter is adapting the canary tests so they run on an ongoing basis.

The researchers are now working with about a dozen providers, and their planted secrets were still being accessed this week.

Is OpenRouter safe? The short answer has two halves. OpenRouter says nobody broke into its own systems, but prompts it handed on to outside AI providers were copied, read and in some cases used to try to break into cloud accounts. That includes prompts carrying a label that was supposed to mean they were never stored.

If you use a chatbot app, a coding assistant or any tool that sends its requests through OpenRouter, this matters because of what people paste into prompts: passwords, API keys, links to private code. On November 22, OpenRouter said it had removed three providers from its platform and taken away the Zero Data Retention status of 125 endpoints. It also says it has contacted every customer who may be affected.

01

Is OpenRouter safe to use right now?

OpenRouter’s own position is that the findings showed no breach or unauthorized access of OpenRouter itself. The trouble sat one step further down the line.

OpenRouter is a gateway, sometimes called an AI router. An app sends it a prompt, and it forwards that prompt to one of many third-party companies that run AI models, often choosing by price or by which model suits the job. The catch is that every company along that path can see the whole prompt and the whole answer.

So the company’s response has been to cut the path, not patch its own servers. The 125 endpoints that lost their Zero Data Retention status include providers the researchers flagged, plus others pulled as a precaution. Any endpoint removed in the review has to offer assurances and pass validation before it gets the status back.

A small yellow paper bird resting on a laptop keyboard on a dark desk
02

How 2 million planted secrets caught the leaks

The evidence comes from research the security firm Dragos published on November 19, led by its chief research officer, HD Moore. Between July and November, the team sent about two million prompts to widely used AI services, each one carrying bait.

The bait was a set of secrets that had never been used anywhere else:

  • Amazon Web Services access keys tied to empty accounts
  • passwords for services such as SSH, a common way to log in to a server remotely
  • one-off web links and DNS names
  • references to private code repositories

Each secret went into exactly one prompt, sent to a single provider. That is the clever part. A secret that exists in only one place can only be learned in one place, so if anyone ever used it, someone had read that prompt. Security people call these traps canaries, after the birds once carried into coal mines.

The canaries sang. The research found that data from at least three AI providers had been exposed, that resellers on an AI marketplace were viewing and keeping customer prompts through ordinary logging, and that some providers appeared to ignore privacy settings or their own published policies.

03

What thieves did with the stolen prompts

The stolen secrets were put to work fast, in some cases within hours. In the biggest of the three break-ins, the bait was a couple hundred AWS keys, sent to a single provider in several brief bursts in late October. Nothing sent during those bursts escaped the thief.

Some of the stolen keys came from requests the provider had turned away before any AI model ran. To the researchers, that suggests the intruder was copying traffic at the front door, not deeper inside. The bait keys were then used for tens of thousands of AWS API calls, including attempts to add new users so the access would last.

The earliest of the three breaches began in mid-July, and none of the three providers knew until the researchers told them. The bait prompts were processed alongside real customer requests, and the firm says those were almost certainly copied as well.

Not every leak needed a break-in. Some sellers on an AI marketplace listed models at the lowest prices, switched on request logging and collected whatever came through. They looked like individuals on home internet connections. One, on home broadband in France, lifted passwords from two prompts and typed server login attempts by hand.

Then there is an unnamed Chinese AI lab. When bait prompts went to that lab’s models, its systems opened the hidden links in over half of them, including about two-thirds of those marked for no data collection or Zero Data Retention. One of those links later turned up in a chat on Feishu, a Chinese workplace app, and was opened within two minutes from desktops and phones in Beijing and Shanghai.

Moore’s point is blunt: marking a prompt as private does nothing once somebody has already broken into the provider.

An open brass padlock lying on a tangle of network cables
04

Which OpenRouter providers have admitted problems

Three providers have posted public notices so far: Phala, Together AI and Novita. OpenRouter’s public list links to all three, but it does not say they are the three it removed, and it says not every affected provider has gone public.

  • Phala said an attacker took control of one of its inference servers and could see requests while they were being processed. It contained the incident on November 8, and its audit found no evidence that other models or products were hit.
  • Novita blamed a routing setup mistake. Some requests for MiniMax models were sent on to MiniMax’s official API when they should have stayed on Novita’s own servers. The affected window runs from May 28 to November 11, and Novita has suspended the models involved.
  • Together AI said on November 21 that an AI coding agent had posted a credential for a private repository to a public GitHub repository. An outsider used it to download stored prompt samples, searched them for logins to other services and tested what they found.

OpenRouter has also emailed customers about prompts sent to MiniMax, a China-based model provider. The planted links were accessed whether prompts went to MiniMax directly or through three resellers, and OpenRouter has taken the Zero Data Retention label off MiniMax-model endpoints.

05

What this means if your apps run through OpenRouter

For ordinary users, the lesson is about what a privacy label is worth. Zero Data Retention is the setting that is meant to stop a prompt and its answer from being stored. In practice, Moore wrote, flagged prompts were kept, handed to other companies and read by human beings.

The providers’ notices give the clearest advice. Phala says anyone whose requests to its DeepSeek V4 Flash 0731 model came through the gateway between October 21 and November 8 should treat the contents as exposed. Every key, token, password or other secret in those requests should be changed.

Together AI is still reviewing the affected prompts and working out which customers it needs to contact directly, so its list is not final. Novita says requests made directly or through OpenRouter in its window may have been accessed or kept by a third party.

06

What OpenRouter is changing next

OpenRouter says it is borrowing the researchers’ method. It is adapting the canary tests so the same kind of check can run on an ongoing basis, and it is technically verifying network paths to find where requests and responses may be forwarded to third parties.

The providers are making changes too. Together AI is moving to short-lived or narrowly scoped credentials, checking every code submission for leaked credentials and turning on Zero Data Retention by default for new accounts. Novita is pressing MiniMax to delete any data it kept. Phala plans a full root-cause analysis.

The story is not over. The researchers are now working with about a dozen providers, and the bait is still drawing bites. As of this week, months on from the earliest thefts, someone was still trying the planted keys and opening the hidden links.

07

Quick answers

Was OpenRouter hacked?

OpenRouter says no. According to the company, the findings showed no breach or unauthorized access of its own systems. The exposure happened at outside providers that receive prompts from it.

What is Zero Data Retention on OpenRouter?

It is a setting under which prompts and responses are not supposed to be stored. OpenRouter has removed 125 endpoints from eligibility for it after the research showed some flagged prompts were kept and read.

Which OpenRouter providers were affected?

Phala, Novita and Together AI have published notices. OpenRouter has not named the three providers it removed, and it says not all affected providers have posted publicly.

You're reading the quick version.