OpenRouter MiniMax Prompts May Have Been Read: What to Do
TL;DR · the short version
OpenRouter emailed customers that "Requests you sent to MiniMax models may have been read." Per the email, a security researcher placed unique test URLs in prompts sent to MiniMax models from July to October 2026, and a third party accessed those URLs 40 to 130 hours later, whether MiniMax served the request or AtlasCloud, Novita or StreamLake forwarded it to MiniMax. OpenRouter says the issue sits with MiniMax and those forwarding providers, not with OpenRouter, that MiniMax has not yet given its own account, and that it has removed the Zero Data Retention label from these MiniMax-model endpoints. OpenRouter says none of the researcher's AWS test keys were used. There is no public OpenRouter page and no statement from MiniMax or the three providers as of 21 November 2026. Treat prompts and outputs sent to MiniMax models in that window as possibly read, rotate any secret that was in them, and set routing and data policies so sensitive requests only reach providers you have approved.
OpenRouter has told customers that prompts and outputs they sent to MiniMax models may have been read by a third party. In an email seen by PlayCISO, titled "Requests you sent to MiniMax models may have been read", OpenRouter says a security researcher placed unique test URLs in prompts sent to MiniMax models from July to October 2026, and that "a third party accessed those URLs 40 to 130 hours after the requests were sent, whether MiniMax served a request directly or another provider forwarded it to MiniMax." OpenRouter says the issue sits with MiniMax and the providers that forwarded to it, not with OpenRouter. MiniMax has not commented publicly.
Who this is for: security leaders, platform and AI engineering teams, and anyone whose apps or staff call large language models through OpenRouter or another LLM router.
Independent analysis based on OpenRouter's customer email (seen by PlayCISO; the same text was posted on Hacker News on 21 November 2026), OpenRouter's public provider-routing documentation, Dragos's 19 November research post, and reporting by RuntimeWire and Startup Fortune. PlayCISO is not affiliated with any company named. Facts are as reported on 21 November 2026 and may change.
Confirmed vs not yet known
OpenRouter says (customer email): test URLs placed in prompts to MiniMax models from July to October 2026 were accessed by a third party 40 to 130 hours later, both when MiniMax served the request and when another provider forwarded it to MiniMax. "With one exception, the same tests on other models at those providers were not accessed, and none of the AWS test keys were used." AtlasCloud, Novita and StreamLake told OpenRouter they forwarded OpenRouter requests for MiniMax models to MiniMax. OpenRouter has removed the Zero Data Retention label from these MiniMax-model endpoints while it investigates.
Relayed, not first-hand: "MiniMax told StreamLake that content moderation was enabled on its endpoints before October 22 and that it has since enabled Zero Data Retention." This reaches us through OpenRouter, via StreamLake.
Not yet known: who the third party is; whether the access was automated scanning or people reading; how long MiniMax kept the data and where; whether content moderation explains the access; which model the "one exception" involved; how many customers or requests are affected in total. OpenRouter says it has asked MiniMax for its own account and has not received one.
Not public: as of 21 November 2026 there is no OpenRouter blog post or status page for this (the Hacker News poster also noted "There is no public page or blog post"), and no statement from MiniMax, AtlasCloud, Novita or StreamLake that PlayCISO could find. OpenRouter does not call this a breach, and nobody has confirmed one.
What happened
OpenRouter is an LLM router: one API that sends each request to one of many model providers. For a given model there can be several providers. OpenRouter's email says that for MiniMax models, three of those providers, AtlasCloud, Novita and StreamLake, did not run the model themselves but "forwarded OpenRouter requests for MiniMax models to MiniMax." So a customer who picked one of them, or let OpenRouter pick, could still have had the request handled by MiniMax.
A security researcher, whom the email does not name, placed unique test URLs (canaries) in prompts. Each URL was used once, so any later visit to it shows that someone or something read that prompt. OpenRouter says the visits came 40 to 130 hours after the requests, which is long after a response would have been returned. Its conclusion: "This means a third party may have read the prompts and outputs of these requests."
OpenRouter says the same tests on other models at those providers were not accessed, "with one exception", and that "none of the AWS test keys were used." Its email places responsibility with "MiniMax and the providers that forwarded to it, not with OpenRouter."
The Dragos research: related, not confirmed as the same
On 19 November 2026, two days before the email surfaced publicly, Dragos published research by its chief research officer HD Moore, "Three breaches, endless leaks: Your AI supply chain needs canaries." It describes about two million prompts seeded with unique URLs and credentials from July to November 2026. Dragos says an unnamed Chinese AI lab's systems "fetched the secret URLs from more than half of the prompts that we sent to its models between July and October", including "roughly two-thirds of the requests marked as 'no data collection' or Zero Data Retention", that "the scraper typically showed up two or three days after the request", and that "several U.S. hosting providers resell this lab's models, and prompts sent to those providers triggered the exact same scraper."
That matches the shape of OpenRouter's email closely. But Dragos does not name the lab and OpenRouter does not name the researcher, so PlayCISO is not treating them as confirmed to be the same case. Dragos's separate findings of breaches at three hosting providers, where its AWS bait keys were used, are about unnamed providers and are not linked to MiniMax by any source; OpenRouter says the AWS test keys in its case were not used.
Timeline
| Date | Event | Source |
|---|---|---|
| July to October 2026 | Researcher places unique test URLs in prompts sent to MiniMax models; a third party later accesses them 40 to 130 hours after each request | OpenRouter customer email |
| Before 22 October 2026 | Content moderation enabled on MiniMax's endpoints (MiniMax's account, as told to StreamLake and relayed by OpenRouter); MiniMax has since enabled Zero Data Retention | OpenRouter customer email |
| 19 November 2026 | Dragos publishes HD Moore's canary research describing an unnamed Chinese lab whose systems fetched canary URLs days later | Dragos blog |
| 20 November 2026 | Ars Technica covers the Dragos research ("A leaky AI supply chain is exposing secret user data at an alarming rate") | Ars Technica, via Hacker News item 50582794 |
| 20 November 2026 (US Central time) | RuntimeWire reports receiving an OpenRouter notice about MiniMax requests routed through four providers | RuntimeWire |
| 21 November 2026 | The OpenRouter email is posted on Hacker News; another user replies that they received the same email; the poster notes there is no public page | Hacker News item 50585962 |
| 21 November 2026 | OpenRouter has removed the Zero Data Retention label from these MiniMax-model endpoints while it investigates, and is awaiting MiniMax's account | OpenRouter customer email |
Note on dates: the email gives July to October 2026 for the researcher's tests. Customer notices list each customer's own affected requests, and RuntimeWire reports its notice covered requests from 15 October to 12 November, so do not assume your exposure stops at the end of October. Check your own notice and logs.
Useful next, if this was relevant to you
Why this is a supply-chain (subprocessor) problem
When you send a prompt to a router, at least three parties can see it in clear text: the router, the provider it picks, and anyone that provider forwards to. In privacy terms the second and third are subprocessors of your data, and the last one may be a company you never chose and never reviewed. Here, per OpenRouter, choosing AtlasCloud, Novita or StreamLake for a MiniMax model still put the request in front of MiniMax.
The same lesson applies to privacy labels. A "Zero Data Retention" or "no data collection" setting describes a policy at one hop. Dragos puts it bluntly: "a privacy flag only binds the first hop." OpenRouter has removed the ZDR label from these endpoints while it investigates. A label is a claim to verify, not a control you own.
We covered a different router risk, routers that tamper with tool calls, in malicious LLM routers. This case is quieter: nothing in the response changed, and the exposure only shows up because someone planted canaries.
What security leaders should do this week
- Find every path through an LLM router. Inventory which apps, agents, internal tools and teams call OpenRouter or any other router or reseller, including developers' personal keys and coding agents configured with a router base URL. Search code and config for router API hosts and keys.
- Check logs for MiniMax models, July to October 2026 and after. Use the router's activity logs and your own request logs to list every call to a MiniMax model in that period, by provider (MiniMax, AtlasCloud, Novita, StreamLake and any other). Read the OpenRouter notice your account received for its own dates.
- Assume those prompts and outputs were exposed. OpenRouter says a third party "may have read the prompts and outputs." Work out what was in them: customer data, personal data, source code, internal documents, credentials. That decides whether privacy, legal or customer notification reviews are needed.
- Rotate any secret that was ever in a prompt. API keys, database URLs, tokens, passwords and signed URLs pasted into prompts or pulled in by an agent's context should be revoked and reissued, in dependency order. The LLM API key exposure playbook covers the steps.
- Enforce data policies at the router. On OpenRouter, the documented controls are
provider.ignoreandprovider.only(skip or allow provider slugs),allow_fallbacks: false,data_collection: "deny"andzdr: true, per request or account-wide in privacy settings. Blocking one provider does not block providers that forward to it, so restrict by model as well as provider for sensitive workloads, and treat ZDR as a policy to verify rather than a guarantee. - Keep secrets out of prompts by design. Scan prompts and agent context for secrets before they leave your network, give agents short-lived scoped credentials through tools rather than in context, and redact personal data where the task allows.
- Review LLM vendors as subprocessors. For each router and provider, ask who else processes the request, where, for how long, for what purpose (including moderation), and how you will be told when that changes. Add the answers to your vendor and subprocessor records.
- Seed your own canaries. Dragos recommends planting unique canary URLs or credentials in a small share of traffic. It is cheap and it is how this exposure was found.
Put it to work
- AI Subprocessor Impact Lab: practise tracing an AI vendor change to the subprocessors that actually handle your data, and choosing a response.
- AI Supplier Questionnaire: the questions to put to a router or model provider about forwarding, retention and moderation.
- LLM Security Assessment: check your controls for data leaving through LLM calls.
- Secret Rotation Order Planner: plan the order to rotate secrets that may have been in prompts.
FAQ
What did OpenRouter tell customers about MiniMax?
In an email titled "Requests you sent to MiniMax models may have been read", OpenRouter said a security researcher placed unique test URLs in prompts sent to MiniMax models from July to October 2026, and that a third party accessed those URLs 40 to 130 hours after the requests were sent. It said this happened "whether MiniMax served a request directly or another provider forwarded it to MiniMax", and that "a third party may have read the prompts and outputs of these requests." As of 21 November 2026 OpenRouter has not published the notice on a public page.
Was OpenRouter hacked?
OpenRouter says no. Its email says: "Based on what the researcher and these providers have reported, the issue sits with MiniMax and the providers that forwarded to it, not with OpenRouter." That is OpenRouter's own assessment; no independent investigation has been published.
Which providers forwarded OpenRouter requests to MiniMax?
OpenRouter's email says "AtlasCloud, Novita and StreamLake have told us they forwarded OpenRouter requests for MiniMax models to MiniMax." So choosing one of those providers on OpenRouter instead of MiniMax itself did not keep a request away from MiniMax. None of the three has published its own statement that PlayCISO could find.
Has MiniMax said anything?
Not publicly. OpenRouter says "MiniMax told StreamLake that content moderation was enabled on its endpoints before October 22 and that it has since enabled Zero Data Retention", and that OpenRouter has "asked MiniMax for its own account and have not received one." Those are second-hand accounts relayed by OpenRouter, not a MiniMax statement.
Were secrets or AWS keys in the prompts stolen?
OpenRouter says "none of the AWS test keys were used" in the researcher's tests. That is about the researcher's test keys, not your data: the email says a third party may have read prompts and outputs, so treat any secret you put in a MiniMax prompt in that window as exposed and rotate it.
Is this the same as the Dragos canary research by HD Moore?
Not confirmed. On 19 November 2026 Dragos published research by HD Moore that seeded about two million prompts with unique canary URLs and credentials. It says an unnamed Chinese lab's systems "fetched the secret URLs from more than half of the prompts that we sent to its models between July and October", that "the scraper typically showed up two or three days after the request", and that "several U.S. hosting providers resell this lab's models". The details line up with OpenRouter's email, but Dragos does not name the lab and OpenRouter does not name the researcher.
Does Zero Data Retention (ZDR) protect my prompts on an LLM router?
Only as far as every hop honours it. OpenRouter has removed the ZDR label from these MiniMax-model endpoints while it investigates. Dragos's research says "a privacy flag only binds the first hop" and that its scraping included "roughly two-thirds of the requests marked as 'no data collection' or Zero Data Retention." Use ZDR and data-collection settings, but do not treat them as a guarantee: keep secrets out of prompts.
How do I stop OpenRouter routing my requests to MiniMax?
OpenRouter documents per-request provider routing settings: "ignore" (a list of provider slugs to skip), "only" (a list of provider slugs to allow), "allow_fallbacks": false (no backup providers), "data_collection": "deny" (only providers that do not collect user data) and "zdr": true (only ZDR endpoints). The same data and ZDR policies can be set account-wide in OpenRouter's privacy settings. Note that skipping the MiniMax provider does not skip providers that forward to MiniMax, so block the model, not only the provider, if that is your aim.
Sources
- OpenRouter customer email, "Requests you sent to MiniMax models may have been read" (seen by PlayCISO; not published on a public OpenRouter page as of 21 November 2026)
- Hacker News: "Tell HN: OpenRouter prompts to Minimax may have been read" (21 Nov 2026): the same email text, posted by a recipient
- OpenRouter documentation: Provider routing (
only,ignore,allow_fallbacks,data_collection,zdr) - Dragos (HD Moore): Three breaches, endless leaks: Your AI supply chain needs canaries (19 Nov 2026); does not name the lab
- Hacker News discussion of Ars Technica's report on the Dragos research (20 Nov 2026)
- RuntimeWire: OpenRouter warns MiniMax prompts may have been read after test URLs were accessed days later (20 Nov 2026)
- Startup Fortune: OpenRouter warns MiniMax prompts may have been read by a third party (21 Nov 2026)
This article is educational guidance for security and engineering leaders, not legal advice. Check OpenRouter's notice to your account and any statement MiniMax or the forwarding providers publish before acting on the details.