November 22, 2026
OpenRouter Prompt Leak: Check Your Writing Tools in 2026
An incident notice needs request records behind it, not a guess based on the model name.

To see whether the OpenRouter prompt leak may have affected you, check your provider notice and request history. In a November 22 statement, OpenRouter said it had removed three providers and ended Zero Data Retention eligibility for 125 endpoints. The findings didn't indicate a breach of OpenRouter itself. Since the notice concerns downstream providers, an app name alone can't tell you whether your writing was affected.[1]
A writing team may need to find out which drafts went through those providers. The model name in a saved document may not show that. Before sending another revision, ask whoever manages the tool to check the affected request routes and dates.
What the current notices establish
Research published by HD Moore at Dragos on November 19 describes nearly two million prompts seeded between July and November. Researchers tracked whether anyone later used the unique links and credentials. They reported compromised providers, marketplace sellers accessing prompts, and privacy settings that weren't honored.[2]
Ars Technica reported the findings on November 20 and added an update about OpenRouter's notification on November 21. I read the full report in an archive captured on November 22. The coverage helps explain why the issue drew attention, but it doesn't identify every affected customer or document.[3]
OpenRouter's statement doesn't say every prompt was leaked. It says customers who may have been affected were notified, and technical checks are still underway. Restoring an endpoint's ZDR eligibility will require assurances and validation.[1] Findings can change. Keep the notice you received, and don't rely on a screenshot of today's label.
On November 21, Together AI issued a separate notice about stored prompt samples from requests in June and July. It said an exposed repository credential let an unauthorized party download them. Together AI says it has tightened credential controls and made ZDR the default for new accounts.[4] That change doesn't explain what happened to an older request.
The disclosures cover different things. Keep their dates separate rather than treating them as one exposure period. Provider incidents may also have had different causes.
Find the request behind the finished document
Imagine a small design studio preparing a proposal. In this hypothetical example, someone speaks a rough project description and asks an AI tool to clean it up. The exported proposal names the model, but not the company that processed the request.
For this studio, checking the final PDF won't show what was exposed. The administrator needs to review the request records the application kept. Ask for the time, selected provider, model, request identifier, and any vendor notice. Mark missing details. If the provider is unknown, enter "provider unknown" in the spreadsheet instead of guessing from memory.
Start with the projects that matter most. In the fictional proposal, this might include the client's unreleased product name and an internal budget estimate. Make a private list of the types of information you submitted. Don't paste the whole proposal into a new AI chat to ask whether it was sensitive.
Check the request counts. One request might contain a whole document, while many others could be harmless layout questions. Keep the vendor's count if they provide one, but don't treat it as the number of clients exposed.
If you can't link an incident notice to specific records, say so. You may only be able to identify an application used on certain dates, with no routing records available. That gives the person responsible for security a clear lead to investigate.
Check what changed before resuming rewrites
A green setting today doesn't show how requests were routed last month. OpenRouter removed ZDR eligibility from 125 endpoints, so timing matters.[1] Keep notes about the affected period separate from the settings you plan to use next.
Ask the app operator which providers they allow for future work. Find out what happens if your chosen provider is unavailable. The app might reject the request or send it to another provider. Check your app's settings for fallback options. Don't rely on another product's help page.
Then try a harmless test with made-up project details. If the tool shows which route the request took, record it. This can confirm the intended settings are active, but it can't prove a provider never keeps data or that future requests will take the same route.
The Dragos report used controlled canaries to detect later access.[2] Writers shouldn't try to repeat the test by pasting real credentials into a model. If actual secrets appear in a prompt that may be affected, notify the person authorized to handle the incident. Keep those details out of public support posts.
The design studio might resume work through an approved account with a confirmed route. Or it might finish one sensitive proposal without another AI rewrite while the operator investigates. It doesn't need to resend the whole file just to meet a copy-editing deadline.
Voice input is one part of that route
Dictating a draft into an editable field can make it easier to revise. But you still need to check where the text goes next. If your dictation tool uses cloud transcription or AI editing, check those steps and the chatbot that receives the text.
With DictaFlow's desktop workflow, hold a shortcut, speak, then release it to add text in a supported field.[6] You can use it to write a proposal or ask about an incident. It can't audit an OpenRouter request or remove context already sent to another app.
Our privacy policy explains how we handle audio and text, as well as selected text and instructions used with relevant editing features.[5] Assess those practices directly. This article doesn't claim DictaFlow was involved in the reported incidents. Changing input tools wouldn't undo an earlier downstream disclosure.
Next time you contact support, send the application name, approximate request dates, and notice identifier through the vendor's private channel. Ask if it can identify which providers handled those requests. Keep the reply with the affected project's records, especially if the vendor can't recover the route.
DictaFlow used AI and the linked sources to help write this article. The design-studio scenario is invented. We didn't repeat the researchers' experiment or review customer request logs.
Sources
[1] OpenRouter provider incident notice, November 22
[2] Dragos research, November 19
[3] Ars Technica, archived November 22 at 14:14:58 UTC