OpenRouter, a service that gives developers one API for reaching AI models hosted by dozens of different providers, has removed three of those providers from its network and revoked Zero Data Retention (ZDR) eligibility for 125 endpoints. The action followed independent research suggesting that some partner infrastructure was either compromised or quietly ignoring the data-handling promises customers were told applied to their prompts.
In a notice published on November 22, OpenRouter said it had contacted every customer who may have been affected. Importantly, the company stated that its own platform was not breached — the problem sits further down the chain, in the providers OpenRouter routes requests to on a customer’s behalf.
Canary Tokens Catch What Audits Miss
The research behind the disclosure, conducted by security firm Dragos, relied on a decades-old trick given a modern twist. Between July and November 2026, researchers sent roughly two million test prompts through OpenRouter’s network. Each one contained a “canary token” — a unique link or a dummy credential that would quietly phone home if anyone other than the intended model ever opened or used it.
That simple mechanism turned up real evidence of mishandling. Investigators traced compromises at three separate hosting providers, including cases where prompts were copied even after the model had refused to process them. In one striking example, a tracking link embedded in a prompt sent to a United States-based provider surfaced in a corporate chat system in China 24 days later — a sign that the data had been forwarded, stored, or otherwise handled well outside the boundaries customers were told to expect.
Why “Zero Data Retention” Isn’t a Guarantee
Zero Data Retention is meant to assure customers that their prompts and the model’s replies are never stored. But as OpenRouter’s architecture illustrates, routing a request and keeping it confidential are two separate engineering problems. A provider that receives a request isn’t necessarily the only party that touches it — some providers forward traffic to other companies for processing, and each additional hop is a new opportunity for a promise to quietly break down.
This is the same dynamic that has surfaced in other AI-supply-chain incidents this year, including exposed gateways that inadvertently left chat histories readable. The common thread is that customers often can’t see past the first link in the chain, so a privacy guarantee is only as strong as every subsequent party’s willingness to honor it.
A Precaution, Not a Confirmed Mass Breach
OpenRouter has been careful to frame its response as precautionary. The 125 endpoints stripped of ZDR eligibility do not represent 125 confirmed incidents — rather, the company is treating any endpoint it can’t currently vouch for as suspect until it has been re-verified. OpenRouter says it is now examining whether providers across its network forward requests or responses onward to additional third parties, and plans to restore ZDR status only after individually confirming an endpoint’s behavior.
The company has not disclosed how many customers or conversations were actually exposed, and has pointed affected users toward its provider disclosure page for specifics on which companies — reportedly including Novita, Phala, and Together — were named in the findings.
What Affected Customers Should Do Now
- Review exactly what was sent through any provider named in OpenRouter’s notice, including attachments and system prompts.
- Treat any password, API key, or access token that appeared in a flagged prompt as compromised — rotate or revoke it immediately.
- Check downstream access logs for unusual activity tied to any credentials that may have leaked.
- Where possible, restrict sensitive workloads to providers you have independently vetted rather than relying on a platform-wide retention policy.
- Avoid placing secrets, credentials, or sensitive personal data inside prompts at all — retention settings are a backstop, not a substitute for careful input hygiene.
The incident is a reminder that the convenience of a unified AI gateway comes with a tradeoff: visibility. Enterprises adopting multi-provider routing platforms should ask not just whether a provider promises not to retain data, but how that promise is verified — and what happens when, as in this case, it turns out not to hold.
Leave a Reply
You must be logged in to post a comment.