AI prompt privacy became an urgent supply-chain issue on November 22, 2026, when OpenRouter said it had removed three inference providers and stripped 125 endpoints of Zero Data Retention eligibility. The action followed research that planted trackable secrets in nearly two million AI requests and found evidence of provider compromises, prompt exposure, and data-policy failures.

If your team sends source code, customer records, internal documents, or credentials to hosted models, the lesson is immediate. A privacy flag can shape routing policy, but it cannot prove that every processor in the path honored the promise. You need to know where prompts traveled, what they contained, and how to respond when a provider's controls fail.

Key Takeaway: Zero Data Retention is a useful policy control, not a cryptographic guarantee. Treat every external inference path as a real data disclosure boundary.

What OpenRouter disclosed on November 22

OpenRouter's November 22 incident update says researchers found that some inference providers had compromised data or showed evidence that their data policies were not being upheld. OpenRouter said the findings did not show a breach or unauthorized access of OpenRouter itself.

The company reported three immediate actions:

  • off-boarded three providers after its investigations
  • removed 125 endpoints from Zero Data Retention, or ZDR, eligibility
  • notified customers who may have been affected

The 125 removals included endpoints flagged by researchers and others removed as a precaution. OpenRouter also said it was adapting the researchers' tests for continuous use, verifying network paths that may forward requests to third parties, and requiring validation before restoring ZDR eligibility.

That distinction matters. A router can enforce a policy at the first decision point while an upstream provider, reseller, subcontractor, or compromised system creates a different reality farther along the path.

Dragos published the underlying research on November 19. Its team sent approximately two million prompts containing unique URLs and canary credentials across about five months. The researchers reported evidence that data from at least three providers was exposed, with the earliest incident dating to mid-July.

Key Stat: OpenRouter removed 125 endpoints from ZDR eligibility after research tested nearly two million canary-seeded prompts.

Why AI prompt privacy breaks across routing chains

A hosted AI request may cross more organizations than the customer expects. The application calls a gateway, the gateway selects an endpoint, one provider may forward to another, and connected tools may receive parts of the context. Each hop can create logs, caches, moderation records, traces, or support artifacts.

Transport encryption does not remove that exposure. TLS protects data between endpoints, but each legitimate processor must decrypt the request to perform its role. Once the prompt reaches that system, operational access, logging configuration, storage controls, and incident response determine what happens next.

ZDR has a narrower meaning. OpenRouter's own ZDR explanation defines it as a provider processing a prompt and response without retaining them afterward. It does not mean the prompt stayed inside your network, remained unreadable to the processor, or avoided every external tool.

This creates four separate questions that teams often combine incorrectly:

  1. Routing: Which organizations and systems received the request?
  2. Retention: Did any processor store prompt or response content after inference?
  3. Training: Could the content be used to improve a model?
  4. Access: Could staff, attackers, or subcontractors read stored or in-flight content?

A “no training” promise does not answer the retention question. A ZDR label does not prove that no compromised administrator copied data while it was being processed. Regional routing does not tell you whether a connected tool stored part of the same request.

Hexon's earlier analysis of malicious LLM routers focused on intermediaries that can alter responses or steal credentials in transit. The new disclosure exposes a different failure mode: even a service selected under a privacy policy can leak or mishandle content when the provider chain is unclear or compromised.

Editorial illustration visualizing ai prompt privacy starts with data minimization in an enterprise cybersecurity context

AI prompt privacy starts with data minimization

The safest secret in a prompt is the one that never entered it. Before changing providers or adding another policy switch, reduce the data your AI workflow sends outside your control.

Classify common prompt content into practical tiers:

  • Public: published documentation, public code, and approved marketing copy
  • Internal: routine operating information with limited business impact
  • Confidential: customer data, nonpublic source code, contracts, and security findings
  • Restricted: passwords, private keys, authentication tokens, health data, regulated records, and merger information

Public content may be appropriate for a broad provider pool. Confidential content should use a small, reviewed allowlist with explicit contractual and technical controls. Restricted secrets should normally be blocked or replaced with references, synthetic values, or narrowly scoped retrieval results.

Use a data-loss-prevention control before the model call, not only a warning in the user interface. Detect common secret formats, high-risk identifiers, and prohibited document classes. Redact or block them before they enter the prompt, including hidden system context, retrieved documents, tool output, attachments, and conversation history.

The same rule applies to coding agents. Do not assume a developer's visible message is the full request. The agent may automatically include environment details, repository files, command output, error logs, or previous turns. Review the assembled payload and default context rules, not just the chat box.

Common Mistake: Telling employees not to paste secrets while an automated agent quietly adds configuration files, logs, or retrieved records to the same request.

The shadow AI security guide explains how unmanaged consumer tools create similar visibility gaps. The control objective is the same: approved workflows, enforceable data classes, and a route that security and privacy teams can actually inspect.

Map every inference provider and subprocessor

You cannot evaluate AI prompt privacy from a model name alone. The same model may be served by its developer, a cloud platform, a specialist host, or a reseller that forwards traffic again.

Create an inference data-flow inventory for every production AI feature. Record:

  • application and business owner
  • gateway, router, model, and selected endpoint
  • provider and known subprocessors
  • region and data-residency setting
  • retention and training policy
  • logging, tracing, caching, and abuse-monitoring behavior
  • connected tools and retrieval systems
  • data classes permitted in requests
  • contract, incident contact, and notification timeline

Then test the real route. Provider selection can change because of cost optimization, failover, model availability, or account defaults. Capture destination metadata for each request where the service makes it available, and alert when traffic reaches a provider outside the approved set.

Prefer explicit allowlists over “any endpoint with ZDR.” A label can be one condition, but high-risk workloads also need a named provider, region, model, and subprocessor path. Fail closed when the required route is unavailable. Silent fallback to an unreviewed endpoint converts an availability event into a privacy incident.

This is also where procurement becomes a security function. Contracts should define retention, subprocessors, incident notification, deletion, audit evidence, and the conditions under which traffic may be forwarded. A provider that cannot explain the full path should not receive your most sensitive prompts.

Pro Tip: Run a quarterly route test from each production workspace. Compare the observed provider path with the approved architecture and investigate any new hop before expanding use.

How to respond if prompts may have leaked

Start with the provider's affected dates, endpoints, models, and customer notices. Do not assume every request through the gateway was exposed, and do not assume an absence of direct evidence means every request was safe.

Build an exposure set from your own records:

  1. identify requests sent through affected providers and time windows
  2. recover request metadata and content only from authorized logs
  3. classify the information that appeared in prompts, responses, tools, and attachments
  4. list credentials, personal data, customer records, source code, and regulated material
  5. assign an owner and remediation action to each exposed data class

Rotate any live credential that entered an affected request. That includes cloud keys, repository tokens, database credentials, OAuth tokens, signing material, webhook secrets, and service-account keys. Search for downstream use before and after rotation, because a stolen credential may have created new users, keys, sessions, or workloads.

The fake AI reseller response guide provides a focused workflow for credential theft around discount access services. Apply the same evidence discipline here: revoke first where the blast radius is high, preserve logs, identify unauthorized use, and rebuild trust from a known-good identity.

For personal or regulated data, involve privacy and legal teams early. Determine whether contractual, statutory, or customer notification duties apply. Keep confirmed exposure separate from possible exposure, and document the basis for each conclusion.

Finally, replace or restrict the affected route. A credential rotation repairs one consequence, not the data path that caused it. Move sensitive workloads to an approved endpoint, direct provider relationship, dedicated deployment, or self-hosted environment according to risk and operational needs.

Editorial illustration visualizing how to verify retention controls instead of trusting labels in an enterprise cybersecurity context

How to verify retention controls instead of trusting labels

No external test can prove a negative with perfect certainty, but teams can demand stronger evidence than a marketing page.

Begin with documentation and contract review. Confirm the exact endpoint-level policy, exceptions for abuse or legal review, metadata retention, tool behavior, and subprocessor list. Ask how the provider prevents ordinary application logs, traces, crash dumps, support tools, and backups from retaining payload content.

Then add technical validation:

  • send synthetic canary URLs or credentials that have no production privilege
  • use unique markers per provider and route
  • monitor for later access from unexpected networks
  • verify that failover and error paths preserve the required policy
  • test whether application, gateway, and observability logs capture full payloads
  • review destination metadata and provider changes over time

Canaries are detection controls, not permission to place real secrets in prompts. They can show that a marker was accessed later, but they may not identify every reader or prove that untriggered content was deleted.

Independent assessments should examine implementation, not only written policy. Useful evidence includes audit scope, retention configuration, access reviews, subprocessor controls, breach history, deletion testing, and incident exercises. Ask when the evidence was collected and whether it covers the exact service and endpoint you use.

Key Takeaway: A trustworthy retention program combines contract terms, endpoint restrictions, data minimization, observed routing, synthetic testing, and an incident plan. No single badge replaces the set.

A 48-hour AI prompt privacy action plan

Security, platform, legal, and AI product teams can turn this disclosure into a bounded review.

During the first 12 hours

  • identify OpenRouter workspaces, affected endpoints, and customer notices
  • freeze new sensitive workloads on unverified provider routes
  • search approved logs for affected request metadata
  • rotate high-impact credentials known to have entered prompts
  • preserve provider notices, routing records, and relevant access logs

During hours 12 through 24

  • classify potentially exposed prompt content
  • hunt for unauthorized credential use and new persistence
  • review privacy, customer, and contractual notification duties
  • enforce provider and model allowlists for sensitive workloads
  • block restricted data classes before inference

During hours 24 through 48

  • document every production inference route and subprocessor
  • test failover behavior and policy enforcement
  • deploy synthetic canaries on approved test traffic
  • assign owners to contract, architecture, and monitoring gaps
  • define evidence required before any removed endpoint is trusted again

Completion should be measurable: affected requests are scoped, exposed credentials are rotated, unauthorized activity is investigated, sensitive routes are restricted, and every exception has an owner and deadline.

OpenRouter's action is significant because it converts research into visible provider consequences. The larger lesson is not that every routed AI request is unsafe. It is that AI prompt privacy must be engineered across the entire inference chain, then verified continuously. Minimize what leaves your boundary, know every processor, restrict routing, and prepare to treat a broken retention promise as a real security incident.