OpenRouter Pulls 125 ZDR Endpoints. Were Your Prompts Private?

November 22, 2026

OpenRouter removed 125 endpoints from zero-data-retention eligibility. Learn what the disclosures prove, what is unknown, and how to review your AI request routes.

OpenRouter Pulls 125 ZDR Endpoints. Were Your Prompts Private?

OpenRouter says it removed 125 endpoints from zero-data-retention eligibility and offboarded three providers after research into inference-data handling. That does not mean all 125 leaked: some removals were precautionary, and the findings showed no breach of OpenRouter itself. The practical question is which organizations handled your requests—not just which model you selected. OpenRouter's November 22 notice.

By OpenVeil Editorial. Research cutoff: November 22, 2026. This is a source-based analysis, not independent incident forensics. The decision worksheet below is our own framework.

What Is Confirmed

On November 19, Dragos described sending approximately two million instrumented prompts between July and November. Unique links and controlled credentials let researchers associate later activity with particular requests. Dragos reports evidence of compromises at at least three providers, reseller access to prompts, and data-handling behavior inconsistent with privacy settings. Those are different findings, not interchangeable descriptions of one breach. Dragos research.

OpenRouter reports customer notifications and further checks of onward routing. Its linked provider statements are not a named list of the three offboarded providers. Official response.

This is for people using hosted AI, developers choosing inference routes, and teams deciding what an assistant should be allowed to read. It is not a finding that every hosted conversation was exposed.

Two Million Test Prompts Are Not Two Million Victims

The study's scale describes research traffic, not a measured count of affected people. Dragos says some planted credentials were used for unauthorized access attempts; in other cases the evidence was later fetching of secret links. The researchers also found that exposure could follow a particular seller rather than every route serving the same model. Method and observations.

A request that produces a working answer is not a privacy audit. Neither is a model's reassurance about what happened to your prompt: the model's response is not an authoritative account of the surrounding infrastructure.

The useful unit of review is a request route with a date and configuration. Treating an entire model family as either safe or compromised loses the information needed to make a defensible decision.

The Provider Notices Describe Different Problems

Together AI: stored samples were downloaded. Together's November 21 notice says an AI coding agent committed a credential to a public GitHub repository. An unauthorized party used it to download prompt samples stored in a private Hugging Face repository, then checked credentials found inside them. The samples involved Minimax M3, GLM 5.2 and Kimi 2.6, primarily on June 24 and July 4, 8 and 14. The company is still reviewing affected contents and says it has not identified additional leakage beyond this incident. It recommends enabling ZDR and says that setting is now the default for new accounts. Together's incident notice.

Novita: requests went to another processor. Novita's November 22 notice says a routing configuration forwarded some MiniMax requests to MiniMax's official API instead of Novita-hosted inference. Its review covers May 28–November 11, including direct requests and OpenRouter traffic through Novita. Novita reports access to test URLs or hostnames, but no observed use of the researcher's AWS test keys. It says the pattern suggests automated content review and that it is seeking confirmation from MiniMax about handling and retention. That assessment is not a confirmed account of every access. Novita's privacy notice.

Phala: a separately disclosed processing-window incident. We already published a route-specific Phala request-review guide. Use its linked primary report for that incident's scope; do not substitute another provider's dates or model list.

Combining these notices into a single claim that every provider stored, trained on and leaked everything would be inaccurate. The appropriate response depends on the particular failure.

What Is Still Unclear

The public notices do not resolve every request's path, or establish whether Dragos's three compromised companies are the three OpenRouter offboarded. OpenRouter's disclosure.

Together's content review remains ongoing. Novita's explanation of automated review is still being checked. A link fetch alone does not establish who read the full text, whether it trained a model, or whether a usable credential was stolen. Together, Novita.

The limits cut both ways: incomplete evidence is not proof of universal exposure, but missing evidence is not a certificate of safety.

ZDR, Training Opt-Out and Local History Answer Different Questions

Our practical distinction is between three boundaries:

Boundary Question to ask What another control cannot prove
Active processing Which organizations receive the request and its attachments? A local history setting does not keep an outgoing request on your device.
Stored copies Which histories, logs, samples or backups can remain, and for how long? A training opt-out is not by itself a deletion instruction.
Permitted use Can the data be used for model improvement, review or another purpose? A retention label is not a complete statement of every processing purpose.

This is a checklist, not a claim that every provider maintains every kind of copy. Ask for evidence that applies to the service, route and account settings you actually used.

For example, deleting a conversation from a visible history can resolve a local visibility concern without establishing what happened during inference. Conversely, choosing a restricted inference route does not secure an unlocked browser profile containing old conversations. These are separate problems worth solving separately.

Our training-versus-retention explainer and Hugging Face routing guide expand those distinctions.

What to Check Before Sending Another Sensitive Prompt

Start with evidence you already have. Do not paste the sensitive prompt into a second assistant to ask whether it was exposed.

  1. Match the notice to your use. Record the application, date range, exact model version and known provider. A provider preference today is not necessarily evidence of last month's route.
  2. Ask the app operator about intermediaries. If you used someone else's interface, ask who handled inference and whether a relevant notice applies. The visible brand is not a complete processor list.
  3. Classify the contents locally. Separate ordinary prose from personal records, internal documents and working credentials. Keep the review itself somewhere appropriate for that information.
  4. Act on applicable secret exposure. Follow the relevant incident guidance and your organization's process. Replace affected credentials through their issuer; deleting a chat does not revoke a key. Review access records where available.
  5. Set a smaller future input boundary. Remove identifiers, use placeholders, and avoid giving an agent access to files or secrets it does not need.

Novita explicitly recommends revoking or rotating live credentials included in affected requests as a precaution. That is more useful than rotating every account merely because a model name appeared in a headline. Novita's customer guidance.

For a purchasing review, ask these questions before choosing a plan: Can you identify the actual processors? What happens if the preferred route is unavailable? Do retention settings apply to attachments as well as text? What account records remain separate? Can you obtain incident notifications and request-level evidence? A cheap token price does not answer any of them.

Where OpenVeil Fits—and Where It Does Not

OpenVeil can help with a narrower, useful goal: doing everyday AI writing and analysis without creating a normal server-side private-chat history record. Private-chat history stays in your browser. OpenVeil does not use your prompts, uploads or outputs to train foundation models. Account, billing and operational records are separate from that chat-history boundary. See the OpenVeil privacy policy.

OpenVeil still requires active processing by OpenVeil and necessary providers. Browser-local history is not fully offline inference, anonymity, zero logs, or a guarantee against an upstream incident. This article does not establish whether any particular OpenVeil request was affected or unaffected by these disclosures. Buying a different workspace does not recover or erase information already sent elsewhere.

Try a low-sensitivity task instead of uploading your incident records: ask OpenVeil to turn a fictional company's three-boundary worksheet into a one-page AI-use policy. Supply invented names and placeholders, not real keys, customer records or private request logs. You can evaluate the writing workflow and browser-local history before deciding whether the service fits.

Adults 18+ can create an account for the one-time ten-action, no-card preview, or choose OpenVeil Plus for $10/month with 10,000 monthly credits. Complete paid checkout through Credits & billing after signup. Credits are finite, and video requires a paid plan. Clearing browser data, changing browsers or switching devices can remove or hide local history; local does not mean backed up everywhere.

Frequently Asked Questions

Did OpenRouter itself get breached?

The company says these findings showed no breach of OpenRouter itself. Official statement.

Does removing ZDR eligibility prove an endpoint leaked?

No. Some removals were precautionary. OpenRouter.

Does this prove my prompts were used for training?

No. Access, retention, onward forwarding and training are distinct claims. Establishing one does not establish all the others.

Is browser-local chat history enough for highly sensitive work?

Not on its own. Decide whether hosted processing is acceptable before sending the material. If your requirement is that the content never leave your controlled environment, a hosted assistant is not the right boundary merely because history is local.

Try OpenVeil Plus for $10/month

For everyday AI work where browser-local private-chat history is the feature you want, sign up for OpenVeil Plus at $10/month, with 10,000 monthly credits. Use it to draft a redacted policy, summarize non-sensitive notes or improve ordinary writing—not as a breach-remediation service.

Want to check the workflow first? Try the ten-action preview without a card. OpenVeil is for adults 18+. Paid checkout follows signup in Credits & billing; video is paid-only. Necessary provider processing still applies.

Sources

Current reporting and discussion helped surface the story; factual incident claims above are attributed to the original research and vendor statements. The comparison, review checklist and fictional trial task are OpenVeil Editorial analysis, not vendor incident findings.

When privacy, account control, uploads, and search matter, OpenVeil gives you a private AI workspace designed for that job.