# OpenRouter Removed 125 ZDR Endpoints: How to Check Your Routes

Find out whether an old request used an affected provider and test what happens when your application can no longer find an approved endpoint.

By Paras Madan | 2026-11-22


OpenRouter said on November 22 that it removed 125 inference endpoints from zero data retention eligibility, removed three providers from its platform and notified potentially affected customers. The [disclosure](https://openrouter.ai/blog/announcements/data-processing-incidents-on-inference-providers/) says researchers found problems at specific inference providers and did not find a breach or unauthorized access at OpenRouter itself. Some of the 125 endpoints were removed as a precaution. The number is not a count of confirmed breaches.

There are two separate questions if you use OpenRouter. Did earlier requests go to a route named in an incident notice? And can your application still serve new requests without falling back to a route your team has not approved?

## Start with the provider and request date

Look for a notification from OpenRouter or the relevant provider, then match its affected model, provider and time window against records your team already holds. The same model ID can send requests to more than one inference provider. An application log that shows only the model cannot tell you which one processed an individual request.

If you have an old generation ID, OpenRouter's [generation metadata reference](https://openrouter.ai/docs/api/api-reference/generations/get-generation) shows a `provider_name` field and request timestamp. Use those alongside any provider or endpoint details in the notice. A record of which provider handled a generation is useful for matching, but it does not independently prove how that provider handled the content. Do not paste an old prompt into a support ticket to demonstrate it was sensitive.

[Novita's notice](https://novita.ai/blog/minimax-requests-privacy-notice/) says some MiniMax requests were forwarded to MiniMax's official API instead of being processed on Novita infrastructure. It identifies May 28 through November 11 for potentially affected requests, including traffic routed to Novita via OpenRouter. [Phala's separate preliminary report](https://phala.com/posts/phala-november-preliminary-incident-report) advises treating contents of DeepSeek V4 Flash 0731 requests routed to Phala from October 21 at 00:00 UTC to November 8 at 05:35 UTC as exposed. A notice about one of these routes does not establish the status of another.

## Handle a matching request before changing traffic

If a request falls inside the applicable scope and contained a live credential, follow the provider's guidance and your incident process for revocation or rotation and account review. A new routing rule does not undo old traffic. If the scope is unclear, preserve existing generation IDs and timestamps and ask the provider or your incident lead to reconcile them rather than assuming the request was safe or exposed.

## Verify that the new route still meets your policy

OpenRouter's [ZDR guide](https://openrouter.ai/blog/insights/zero-data-retention/) explains that `provider.zdr: true` requires a currently eligible endpoint. Account privacy settings and workspace guardrails can enforce the same requirement. `data_collection: "deny"` filters according to a separate data collection policy. Neither setting keeps the prompt inside your network or governs your own logs or third-party tools.

Pick one model and use a synthetic prompt in a test environment with the same approved provider and retention rules as production. Record the provider from generation metadata. Then check the case where none of your approved routes can serve the request. [OpenRouter documents a 404](https://openrouter.ai/blog/announcements/guardrails/) for a request disallowed by routing guardrails. Your application should report that the compliant route is unavailable rather than strip the restriction and retry. If you need a backup, test it under the same policy before making it part of your failover.

A successful test tells you that your app followed its current routing configuration. It does not establish how an outside provider implemented its policy or what happened to historical requests. Keep the routing test and the incident review as separate tasks.

I'm Paras. I help AI teams write about technical changes and the choices developers have to make because of them.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
