explainx.ai350K+ learn from us
TrendingAI News TodayExplainx for Business
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

follow on google

Add explainx.ai as a preferred source

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

community

Join the community

learn

mind: share how you thinkpathways — start freeworkshopsbootcampscoursescompare Explainxcertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcpestacktoolsmdx readeragentsllmsdesignsdictionarypeopleagi trackerfelony benchranks

company

aboutvisionmissionteaminstructorsteach on explainxpartnershipscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportcontactprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

explainx.ai

On this page

  • TL;DR
  • What OpenRouter told customers
  • Why the delay matters more than the brand name
  • Zero Data Retention versus Safety Retention
  • How this fits the wider AI supply chain story
  • What developers should do this week
  • Is OpenRouter still worth using?
  • What we do not know yet
  • Is OpenRouter down or broken?
  • Questions paying teams should ask procurement
  • MiniMax-specific background for new readers
  • A simple rotation checklist (30 minutes)
  • Related reading
← Back to blog

explainx / blog

OpenRouter Warns MiniMax Prompts May Have Been Read: ZDR Labels Removed

OpenRouter, MiniMax, AI Security, Data Privacy, AI Routers

Part of AI Tools and Apps

OpenRouter emailed customers that MiniMax prompts from July–October 2026 may have been read by a third party. What the Dragos canary study found, which hosts forwarded traffic, and how to rotate keys.

Nov 21, 2026·9 min read·Yash Thakker
add explainx.ai
go deep
OpenRouter Warns MiniMax Prompts May Have Been Read: ZDR Labels Removed

OpenRouter began emailing paying customers on November 21, 2026 with a blunt subject line: requests you sent to MiniMax models may have been read. The message ties a third-party access pattern to prompts routed to MiniMax between July and October 2026, names resellers that forwarded traffic to the Chinese lab, and says OpenRouter stripped Zero Data Retention (ZDR) badges from the affected MiniMax endpoints while it investigates. There is still no public blog post on openrouter.ai/blog; the notice lives in inboxes first.

This is the customer-facing half of a story Dragos published two days earlier: HD Moore seeded roughly two million prompts with canary tokens and watched where secrets resurfaced. Builders who routed agents through OpenRouter for cheap MiniMax inference now need a rotation plan, not a headline recap.

TL;DR

table · 2 cols
QuestionAnswer (Nov 21, 2026)
What happened?Customer email: test URLs in MiniMax prompts were opened 40–130 hours after inference
Who might have read prompts?OpenRouter blames MiniMax and hosts that forwarded to it, not its own core platform
What changed on OpenRouter?ZDR labels removed from MiniMax-model endpoints pending investigation
Resellers namedAtlasCloud, Novita, StreamLake said they forward MiniMax traffic to MiniMax
Research backdropDragos canary study (Nov 19) on prompt theft and ignored privacy flags
What to doRotate keys, audit July–Oct MiniMax usage, avoid sensitive prompts on marketplace routes
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

What OpenRouter told customers

Several developers posted the full text of OpenRouter's email on Hacker News on November 21, 2026. The message is personalized: it lists how many requests your account sent to MiniMax models, which upstream hosts carried them, and the UTC date range.

The facts OpenRouter attributes to the unnamed security researcher are specific:

  • Unique test URLs were embedded in prompts aimed at MiniMax models from July through October 2026.
  • A third party fetched those URLs 40 to 130 hours after the requests completed.
  • The pattern held whether MiniMax answered directly or another provider forwarded the call to MiniMax.
  • Parallel tests on other models at the same hosts were mostly untouched, and none of the planted AWS keys were used (per the email text reporters shared).

OpenRouter's conclusion in the email is explicit: based on provider reports, the problem sits with MiniMax and the companies that forwarded OpenRouter traffic to it, not with OpenRouter's own routing core. That wording matters for legal and compliance teams, but it does not erase the prompts that already left your environment.

While investigating, OpenRouter says it removed the Zero Data Retention label from the MiniMax-model endpoints in question. That is a product change paying users can see in the model catalog today even without the email.

Providers in the forwarding chain

The email names three hosts that told OpenRouter they forward MiniMax listings to MiniMax itself:

  1. AtlasCloud
  2. Novita
  3. StreamLake

It also relays MiniMax's statement to StreamLake: content moderation was enabled on MiniMax endpoints before October 22, and MiniMax has since enabled Zero Data Retention on its side. OpenRouter says it asked MiniMax for its own account of events and had not received one as of the email send.

None of those sentences appear on a public OpenRouter status page yet. For how OpenRouter describes retention tiers when they are honored, see its data collection documentation.

Why the delay matters more than the brand name

Dragos did not name the Chinese lab in its November 19 post. Reporters covering the OpenRouter email note the same July–October window and multi-day lag between prompt and URL fetch. A real-time content filter might hit a link within seconds. Days fit stored prompts being opened later.

Moore wrote that prompts sent to U.S. hosting providers reselling the lab's models still triggered the same scraper, meaning traffic marketed as domestic could flow back to the lab. OpenRouter's email about forwarded MiniMax routes rhymes with that paragraph even though neither document cross-links.

If you already read our Dragos summary, treat this email as action items for OpenRouter accounts, not a second research paper.

Zero Data Retention versus Safety Retention

Vault door with a key hanging beside it, suggesting stored secrets at riskVault door with a key hanging beside it, suggesting stored secrets at risk

OpenRouter's docs draw a line builders often miss when they tick a privacy box:

table · 2 cols
SettingWhat OpenRouter says it means
Zero Data Retention (ZDR)Prompts and completions should not be stored after processing, aside from required safety paths
Safety RetentionProvider may retain prompts to prevent abuse even when you asked for no training use

The MiniMax email concerns ZDR labels on specific endpoints. Removing the badge does not prove every host was malicious; it signals OpenRouter will not promise ZDR on those routes until the investigation closes.

Moore's blunt summary still applies: a privacy flag only binds the first hop. Our semantic caching and router cost guide explains why multi-hop inference is cheap and why cheapness can hide extra readers.

How this fits the wider AI supply chain story

The MiniMax email is one tile in a busier November:

  • Dragos (Nov 19): three provider breaches with stolen canary AWS keys, marketplace resellers logging prompts, and ignored ZDR paths, from ~two million seeded prompts.
  • Railway's October secret leak: attackers hunted AI API keys in hosting env vars; OpenRouter spend was in the reimbursement table.
  • Google GTIG on LLMjacking: stolen cloud AI keys turned into mixer services reselling access.

OpenRouter is also the router behind Hugging Face Inference Providers since November 16. That partnership does not cause this incident, but it shows how central the router is for everyday model access.

For background on why developers flocked to MiniMax on routers, see our MiniMax M3.1 Flash pricing post and the OpenRouter default routing change from October.

What developers should do this week

1. Read your email, then read your logs

OpenRouter's message includes per-account request counts and hosts. Match that window against:

  • CI jobs that sent repo context through MiniMax
  • Agent harnesses with long system prompts holding credentials
  • Experiments that assumed ZDR because the catalog said so

2. Rotate secrets that lived in prompts

Green key enclosed in a capsule beside a closed folderGreen key enclosed in a capsule beside a closed folder

If a prompt could contain API keys, database URLs, customer PII, or unreleased code, rotate those assets even when the email says AWS canaries were untouched. Canary keys prove read access, not the only thing worth stealing.

3. Pick routes deliberately

Moore's checklist from Dragos still fits:

  • Self-host or go direct to a first-tier provider for sensitive work.
  • On routers, choose upstream providers instead of accepting the cheapest pool.
  • Seed your own canary tokens in agent context and treat firings as incidents.

For agent pipelines that auto-call tools, add guardrails before outbound network calls. AgentBeam can stop agents from exfiltrating secrets once a model or router misbehaves.

4. Watch for a public post-mortem

As of this publish time, explainx.ai sees email plus docs, not a blog post. Check openrouter.ai/blog, your OpenRouter dashboard notices, and whether ZDR badges return on MiniMax listings.

Is OpenRouter still worth using?

For many builders, yes, with eyes open. OpenRouter's value is one key, many models, as we covered in the enterprise OpenRouter guide. This incident does not delete that convenience. It does mean model badges are claims you must verify, especially on discounted marketplace listings that may forward overseas.

Treat the November 21 email like a billing anomaly alert: investigate, rotate, document, then decide whether MiniMax routes belong in production at all.

What we do not know yet

Gaps worth stating plainly:

  • OpenRouter has not published the researcher name; reporters match the details to Dragos, but neither party has confirmed that link on the record.
  • MiniMax has not issued a statement OpenRouter could quote directly, only a relay through StreamLake about moderation and ZDR.
  • Some outlets report longer windows (for example into November) for specific accounts; reconcile against your email, not forum anecdotes.
  • Removing ZDR labels is not the same as proving every prompt was read; it is OpenRouter pausing a promise until evidence is sorted.

Is OpenRouter down or broken?

Some searchers confuse product incidents with availability. As of November 21, 2026, explainx.ai saw no OpenRouter status-page outage tied to this email. The warning is about historical MiniMax routes and ZDR label changes, not a platform-wide failure. If API calls fail today, check OpenRouter status separately from this privacy story.

Questions paying teams should ask procurement

Security review templates rarely mention inference marketplaces. After the MiniMax email, ask vendors:

  1. Which upstream host runs each model SKU in production, not only the router brand.
  2. Whether ZDR is contractual or a catalog badge that can disappear overnight.
  3. Whether reseller forwards to non-US entities are disclosed in your DPA.
  4. How agent frameworks log prompts when they fan out to five models for one task.

Those questions pair with our token fraud and free-trial abuse coverage: cheap inference often funds itself with data you did not price in.

MiniMax-specific background for new readers

MiniMax is a Chinese lab whose models became router favorites when Space Bunny topped usage charts. Low per-token pricing makes MiniMax attractive for batch translation, agent loops, and coding harnesses that burn millions of tokens. That economics story does not change because of one email, but it explains why so many accounts touched MiniMax in July through October even if they never chose the lab directly.

A simple rotation checklist (30 minutes)

table · 2 cols
StepAction
1Export OpenRouter usage or read the email counts for MiniMax models
2List secrets that appeared in those prompts (keys, URLs, customer rows)
3Rotate each secret at the source system, not only on OpenRouter
4Create a new OpenRouter key with spend caps for remaining tests
5Switch sensitive workloads to direct Anthropic or OpenAI routes temporarily
6File a ticket if your compliance team needs OpenRouter's written confirmation

If you route through Bedrock tokenomics dashboards, add a FinOps note so finance sees why November spend shifts providers.

Related reading

  • Dragos canary token study (Nov 19)
  • Railway breach and OpenRouter reimbursement map
  • Hugging Face routes chat models through OpenRouter
  • OpenRouter Armor tool-call safety beta
  • November 2026 cyberattacks tracker
Spotted something out of date? Let us know.
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

  • ✓ Free
  • ✓ One email a day, no spam
  • ✓ Unsubscribe anytime
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

View Yash Thakker in People in AI →

Related posts

Nov 19, 2026

Dragos Seeded 2 Million AI Prompts With Canaries: Three Provider Breaches and Ignored ZDR Flags

Dragos chief research officer HD Moore published results from nearly two million prompts seeded with unique canary tokens. The tokens were used by attackers at three hosting providers, by marketplace resellers, and by systems that were never meant to see ZDR-flagged prompts. We summarize the findings, the caveats and a protection checklist.

Oct 31, 2026

VOLTA Infostealer Uses an LLM to Rank Victims by What They Are Worth

Cisco Talos documented VOLTA-AUTODOXER, a Windows infostealer that steals everything, then asks an LLM to write a target assessment titled "What they are worth". Here is what the report confirms, how the model fallback chain works, and what defenders and ordinary users should do.

Nov 21, 2026

AWS: Cairn AI Agents Found a MONAI Label Zero-Day in 11 Minutes

On November 20, 2026, Amazon Threat Intelligence published an investigation into a China-linked actor who left a public Cairn autonomous pentest server online. Cairn agents hunted shell access at East Asian universities, failed for hours against hardened Redis, then found and exploited an unpatched path traversal flaw in MONAI Label in about eleven minutes. AWS notified maintainers; GitHub advisory GHSA-qc22-rmgc-vg94 lists no shipped patch yet.