OpenRouter MiniMax Prompts May Have Been Read: ZDR and Rotation
Separate Novita’s confirmed MiniMax forwarding from the OpenRouter email reproduced by recipients. Audit the correct time windows, retrieve request metadata, revoke exposed credentials, and review private-code disclosure. Understand current ZDR controls and compare access routes without treating a gateway change as proof of safety.
Contents

OpenRouter’s “Requests you sent to MiniMax models may have been read” notice, reproduced by a recipient, calls for a request audit and credential review. Developers who used MiniMax through OpenRouter or Novita should check the reported windows before sending more sensitive content. Pause those workflows, preserve existing records, and revoke live secrets that appeared in affected prompts or outputs.
What do the two notices establish?
As of 2026-11-22, Novita has published a first-party confirmation of request forwarding. OpenRouter’s account below comes from an email reproduced by Hacker News recipient dannyw. Keep that attribution separate from Novita’s confirmed findings.
| Source | Relevant window | What it establishes or reports |
|---|---|---|
| Novita’s official privacy notice | 2026-05-28 to 2026-11-11 | Some MiniMax requests reached MiniMax’s official API instead of Novita-hosted infrastructure. The advisory covers direct Novita requests and requests routed through OpenRouter. Novita says it suspended affected models. |
| Recipient-reproduced OpenRouter email | Research tests during July–October 2026; separate, personalized account intervals | Unique test URLs were accessed 40–130 hours later, including direct and forwarded MiniMax requests. The email says related endpoints lost their ZDR label. This is not a count of proven reads of customers’ requests. |
Novita attributes the forwarding to a configuration issue that broadened a route intended for specific customers. The email separately names AtlasCloud, Novita, and StreamLake as providers that told OpenRouter they forwarded MiniMax requests. Those statements have different provenance; Novita’s notice independently confirms only its own forwarding.
URL access shows that the unique addresses were used after inference. It does not establish who read every prompt or whether review was human. Novita describes automated content review as a possibility still being checked. Both notices report no observed use of the AWS test keys; that finding cannot clear every customer credential.
The email also relays a claim through StreamLake about moderation before October 22 and subsequent ZDR enablement. Do not treat that secondhand statement as a verified remediation date or a reason to shorten Novita’s window.
Audit requests, credentials, and exposed code
Match requests to the correct window and route
Use the UTC interval in your own OpenRouter notice, plus Novita’s broader advisory wherever Novita handled your traffic. July–October describes the researcher’s tests, not a complete exposure boundary. If your Novita records fall near either date boundary, ask Novita to clarify the applicable timestamps.
Without an account-specific notice, start with all your saved MiniMax requests, rather than guessing an exposure window. Do the same when you cannot confirm whether Novita handled your OpenRouter traffic. Ask OpenRouter support to confirm the affected dates and provider routes for your account. Keep unverified requests open in the audit until you can reconcile them; a missing notice does not establish that they were unaffected.
Build a restricted-access inventory with request ID, timestamp, historical model ID, reported provider, originating job, and the location of saved content. Search all MiniMax IDs your application used, rather than only today’s model name.
OpenRouter’s Logs documentation says logs contain metadata by default. Enabling content logging now cannot recover earlier prompts. Use existing application traces and agent transcripts to reconstruct what was actually sent; do not create a new unrestricted prompt archive.
For an existing generation ID, the generation metadata API retrieves the recorded model and provider without resending the prompt. Run this in Bash, with curl, jq, and an authorized OpenRouter API key. Enter the key at the first, hidden input, then enter the generation ID.
(
set +x
set -euo pipefail
read -rs OPENROUTER_API_KEY && export OPENROUTER_API_KEY
read -rp "Generation ID: " GENERATION_ID
[[ -n "$OPENROUTER_API_KEY" && -n "$GENERATION_ID" ]]
curl --fail --silent --show-error --get \
'https://openrouter.ai/api/v1/generation' \
--header "Authorization: Bearer ${OPENROUTER_API_KEY}" \
--data-urlencode "id=${GENERATION_ID}" \
| jq -e --arg id "$GENERATION_ID" \
'.data | select(.id == $id) | {id, created_at, model, provider_name}'
)
Success means the command returns an object whose id matches your saved generation ID. Compare created_at, model, and provider_name with your inventory. A missing record, failed lookup, or null provider remains unresolved; reconcile local records or ask the account administrator for help.
The reported provider is not proof of the complete processing chain. Given Novita’s confirmed forwarding, a Novita entry cannot, by itself, rule out subsequent forwarding to MiniMax. Record that uncertainty rather than marking the request safe.
Revoke secrets that appeared in transmitted content
Inspect prompts, outputs, tool messages, pasted configuration, and command output available in your saved records. Prioritize administrator, cloud, deployment, and CI credentials. A file’s presence in a repository is not proof that its contents were transmitted; distinguish known disclosure from missing evidence.
Revoke exposed credentials at their issuing service, then update dependent jobs and applications with replacements from your secret store. Do not leave a suspected compromised credential active merely to preserve a routine rotation overlap. OWASP’s secrets-management guidance recommends revocation when secrets may be compromised.
Verify both sides: the old credential is disabled or rejected, and the replacement performs the required authorized operation. AWS’s access-key rotation guidance explicitly includes updating applications and confirming they work with the new key. Creating a replacement alone does not revoke the original.
Review passwords, signed URLs, and temporary sessions using the issuer’s own invalidation mechanisms. For AWS role credentials, assess role-session revocation separately; its scope can include other users of the same role.
Calling MiniMax does not itself establish that a gateway key used only in an authentication header was disclosed. Prioritize secrets present in the affected content, while investigating any independent evidence of account or key compromise. That distinction follows the notices’ stated scope, not a guarantee that header credentials were safe.
Use the AI-agent secret-exposure response guide for service-by-service containment. After containment, review key ownership, expiration, and planned rotation for longer-term governance.
Check misuse and code disclosure separately
Review the credential issuer’s audit logs from the earliest possible disclosure through revocation and follow-up. Investigate unfamiliar source addresses, unexpected operations, new credentials, and access to repositories or secret stores. As AWS’s compromised-credential guidance recommends, confirm suspicious activity with the workload owner before declaring misuse.
For AWS, the default CloudTrail Event history covers only 90 days and only management events. It cannot by itself clear the older portion of this May–November window or every data-access operation. Look for previously collected trail archives and relevant service-access logs, and record gaps explicitly.
Private code needs a disclosure assessment, not just key rotation. Record the files, commits, and customer data actually included, then scan retained copies locally for embedded secrets. Give the repository or data owner that inventory. Ask the processing provider which downstream parties received the content and what deletion it can confirm.
Preserve original evidence with restricted access; use redacted copies for investigation tickets. Novita says it is following up about retained data and deletion, not that deletion has been completed. Do not upload suspect transcripts to another model for analysis.
What does removing the ZDR label change?
Treat the withdrawal reported in the reproduced OpenRouter email as loss of that endpoint’s retention assurance during the investigation. It is neither proof that every request was read nor evidence that previously retained material was deleted. Novita’s separate deletion inquiry remains open.
OpenRouter’s current ZDR definition excludes routine storage after processing, with a stated CSAM exception. Its separate Safety Retention category permits retention for abuse prevention without training. Therefore, a no-training policy alone does not establish zero retention.
For future requests, set zdr: true inside the request’s provider object and select ZDR only in account privacy settings. The latter excludes Safety Retention endpoints. Use provider.only to restrict routing to reviewed provider slugs, following the provider-selection documentation. Do not substitute a guessed provider name for an approved processing path.
Test the configuration with non-sensitive synthetic input and inspect its recorded provider. That checks operation, not downstream compliance. If no acceptable endpoint is available, keep the sensitive job paused rather than relaxing the policy. Review your application’s own logging and enabled tools separately.
Which route can you use while the audit is open?
Choose a route based on the processing arrangement you can approve, not an assumption that a different hostname is safer. Confirm account eligibility and payment options before migrating.
| Route | Requirements and billing | When to choose it |
|---|---|---|
| OpenRouter with reviewed providers | Your account and API key; a client that sends provider controls; text-token charges for the selected endpoint. | Keep an existing integration only when its allowed providers and downstream handling meet your requirements. |
| MiniMax’s own API | A MiniMax account, API key, compatible client, and funded pay-as-you-go usage; text-token billing. | Choose a direct provider relationship. Direct access is not, by itself, evidence that the reported retention issue is resolved. |
| BetterToken | Your own account and API key, a compatible API client, and pay-as-you-go usage. MiniMax-M3.1-Flash is listed as of 2026-11-22; text usage is token-billed. Check current availability and prices. | An alternative access and billing route, subject to the same upstream-retention review. No security advantage is established here. |
| Self-hosted MiniMax open models | Hardware, a supported runtime, access controls, and operational capacity; infrastructure costs rather than a managed API bill. | Consider this when you need to operate the inference environment. Review logs, integrations, and outbound traffic instead of assuming local deployment is automatically private. |
Before resuming confidential traffic, request the actual processing path, applicable retention policy, and any confirmed remediation. A successful API response cannot answer those questions.
Is the Dragos report the same incident?
That connection remains publicly unconfirmed in the sources checked as of 2026-11-22. HD Moore’s November 19 Dragos report describes approximately two million canary-bearing prompts submitted during July–November. A canary is a unique test URL or credential used to detect later access. The report does not name MiniMax, and the reproduced email does not name Dragos or Moore. Similar timing and test methods do not establish identity.
Do not attribute the broader report’s credential-use findings to these MiniMax requests. Likewise, Phala’s November 10 report concerns a separate incident involving Phala-served requests; it is not Novita’s forwarding notice.
Finish this audit with a request inventory, evidence of credential revocation, tested replacements, documented log gaps, and an approved route for the next job. Leave unresolved downstream access or deletion questions open rather than closing them because billing or generation works again.