Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore
Security NewsNews Explainer

OpenRouter Prompt Exposure: Verify Every Inference Provider Hop

OpenRouter prompt exposure raises provider routing and retention questions. Use this checklist to map inference hops and verify endpoint ZDR claims.

By
Brandefense Research Team
Reviewed by
Brandefense Analyst Team
Published
22 November 2026
Reading time
6 min read
Layered AI inference providers with an audit trail and privacy boundary

OpenRouter prompt exposure is a supplier assurance problem for teams that send sensitive data through AI routing services. On November 22, 2026, OpenRouter said it had off-boarded three providers and removed 125 endpoints from Zero Data Retention (ZDR) eligibility. It said the research did not show a breach or unauthorized access to OpenRouter itself.

Dragos reported separate evidence of provider compromises, marketplace sellers viewing prompts through ordinary logging, and retention or forwarding that conflicted with privacy settings. The buyer’s task is to prove where each request travels and who can retain it. The workflow below is our operational interpretation of the two organizations’ public statements.

OpenRouter Prompt Exposure: Which Route Does Your Request Take?

Start with the route used by each application, model and account. Dragos says a request can move through multiple routers and resellers before reaching the inference server. Every intermediary can potentially see the prompt and response, even when the first router displays a privacy control. The provider selected in a user interface may therefore be only one visible part of the processing chain.

Export current routing settings, model identifiers, endpoint identifiers and any provider allowlists from the console or API configuration. Ask the router for the processing chain behind each selected endpoint, including subcontractors and geographic processing locations. Compare that answer with your contract, data-processing agreement and network or application telemetry. Record differences as supplier questions rather than assuming a hidden hop is malicious.

A third-party risk management platform can keep the supplier owner, contractual promise and verification evidence together. The third-party risk management framework also helps distinguish an approved provider from an unreviewed subcontractor. Apply those records to the actual inference route, not just to the company that sends the invoice.

What Should a Buyer Verify First?

Work from data sensitivity to routing evidence, then test the provider’s retention claims. This order limits new sensitive submissions while you investigate existing use. It also makes every exception traceable to an application owner, rather than leaving a broad AI service approval in place.

Inference provider assurance workflow
  1. InventoryExport each application’s router configuration and request log; record model, endpoint ID, privacy setting and prompt data class.
  2. ConstrainIn the application routing configuration, disable sensitive traffic to unverified endpoint IDs and record the change ticket.
  3. MapSave the router’s endpoint-level processor statement and subcontractor list; compare them with the route in the request log.
  4. VerifySave the dated ZDR eligibility record, provider retention schedule and contract clause; check prompt, response, access and deletion logs.
  5. TestSend one harmless canary through an authorized test route; match its alert log to the request log and endpoint ID.
  6. ReassessDiff the saved route configuration and ZDR eligibility record after each change; log the exception owner and review date.

For the inventory, include agent workflows that may copy retrieved documents or secrets into context automatically. Separate production traffic from authorized test traffic so a canary does not enter customer data. Keep a record of which application, prompt category and privacy setting produced each test. Dragos linked its canaries to those fields to make later observations attributable to a route.

Ask each provider for a current retention schedule covering prompts, responses, logs, abuse monitoring, backups and subcontractors. Request a dated explanation of how a ZDR request is enforced at every hop. An account toggle by itself proves only the setting you chose; it does not prove downstream processing. OpenRouter says endpoints removed from its ZDR program need assurances and validation before restoration.

How Should Teams Interpret the Published Evidence?

Dragos sent approximately two million controlled prompts from July through November 2026 and tracked unique canaries. Its report says the observations indicated exposure at at least three AI providers. It also describes resellers that could view prompts through standard marketplace logging without exploiting the router. Those are different failure modes and need different supplier questions.

For one provider route, Dragos observed later use of planted credentials after its test prompts were submitted. For other routes, it observed planted links being accessed well after a single request, including requests marked for privacy controls. A canary shows that a value from a particular test prompt was later accessed; it does not, by itself, identify every intermediary that handled the prompt. Dragos used additional route analysis to investigate that gap.

OpenRouter’s announcement says the research did not show unauthorized access to OpenRouter. Its three off-boarded providers and 125 ZDR eligibility removals are platform actions, not counts of breached customers. The 125 include endpoints removed as a precaution, so the number must not be read as 125 confirmed leaks. OpenRouter said it notified potentially affected customers and was checking network paths for forwarding to third parties.

For context, Brandefense Blackhole recorded 616 dark-chat posts matching OpenRouter during August 24 to November 22, 2026. That is a count of matched posts, not distinct incidents, customers or confirmed prompt exposures. The count should not change an exposure verdict without route and provider evidence. Keep the public notices and your own processing records as the decision basis.

What Evidence Closes a Supplier Exception?

Close an exception only when the supplier names the actual processing entities and explains how each one handles prompts and responses. For a ZDR claim, ask for endpoint-level scope, retention controls, subcontractor terms and a dated validation method. If the chain changes dynamically, require a notification or a route restriction that keeps approved processors in scope.

Keep the verification packet with the application inventory, selected route, provider response, contract clause, test result and review date. Have the application owner document whether sensitive data may resume on that exact endpoint. Reopen the review when routing, pricing or provider eligibility changes. A supply chain threat analysis can help frame the same decision across other delegated services.

Where planted credentials were used in a test, inspect your own cloud audit records for attempted use of real credentials exposed through comparable prompts. Rotate credentials based on verified exposure in your environment and preserve the related logs. Dragos describes attempted access to its controlled accounts, but that observation does not establish compromise of every customer. Treat this as a targeted investigation, with scope set by your routes and data inventory.

Frequently Asked Questions

Did OpenRouter report a breach of its own systems?

No. Its November 22 announcement says the research did not show a breach or unauthorized access to OpenRouter. It reported actions against inference providers and endpoint ZDR eligibility.

Does a ZDR setting prove every inference provider discards the prompt?

No. Verify the endpoint's current eligibility, each processor in the route, retention terms and downstream controls. OpenRouter says removed endpoints need assurances and validation before ZDR eligibility returns.

What does a prompt canary prove?

Later access to a unique, controlled canary links an observation to its test prompt. Dragos says additional route analysis is needed to identify where in the chain the data escaped.

Sources

  1. OpenRouter, Data Processing Incidents on Inference Providers, November 22, 2026
  2. Dragos, Three Breaches, Endless Leaks, November 19, 2026
Share this articleLinkedInXEmail
Brandefense logo

Stay Ahead of Digital Threats.

Take control of your digital security with an exclusive demo of our AI-driven threat intelligence platform.

Open The Full Demo Form →