Skip to content

OpenRouter Warns Prompts Sent to MiniMax Models May Have Been Read

Published

Laura Bede

By Laura Bede

Journalist

Miklos Zoltan

Fact-Checked this

OpenRouter has told customers that prompts they sent to MiniMax’s AI models may have been read by someone else.

The warning went out by email, not as a public notice. A security researcher had hidden unique test links in prompts sent to MiniMax models between July and October. A third party opened those links 40 to 130 hours later.

It didn’t matter whether MiniMax answered a request itself or another host passed it on. OpenRouter names three hosts, AtlasCloud, Novita and StreamLake, that told it they forwarded requests for MiniMax models to MiniMax.

RuntimeWire, a small tech site that says it got the notice too, reported it first, at 03:54 UTC on Saturday, 21 November. About 40 minutes later a Hacker News user posted the full email. When we published, we found no coverage in major outlets, and OpenRouter’s announcements page didn’t mention it.

What OpenRouter Told Customers

The email is titled “Requests you sent to MiniMax models may have been read.” It tells each customer how many requests their account sent to MiniMax models, through which hosts and in what time window.

What the researcher’s tests showed, according to OpenRouter:

  • Links opened days later: the test URLs were accessed 40 to 130 hours after the requests, whether MiniMax served them or a host forwarded them.
  • Only MiniMax: with one exception, the same tests on other models at those hosts weren’t touched. OpenRouter doesn’t say what the exception was.
  • Keys left alone: none of the AWS test keys in the prompts were used.

OpenRouter says the problem “sits with MiniMax and the providers that forwarded to it, not with OpenRouter.” It also passes on MiniMax’s only explanation so far, given to StreamLake rather than to OpenRouter: content moderation was switched on for its endpoints before 22 October, and MiniMax has since enabled Zero Data Retention. OpenRouter says it asked MiniMax for its own account and hasn’t received one.

While it investigates, OpenRouter has removed the Zero Data Retention label from these endpoints. Its conclusion is blunt: a third party “may have read the prompts and outputs of these requests.”

The window may be wider than July to October. RuntimeWire says its own notice covered 80 requests sent through four hosts between 15 October and 12 November, and told it to rotate any secrets in them. That runs three weeks past 22 October, the date MiniMax’s explanation points to.

The Research Behind It

OpenRouter doesn’t name the researcher. The details match research HD Moore published on 19 November as chief research officer of the security company Dragos. From July to November, his team sent about two million prompts with unique links and fake credentials to AI services and watched where they turned up.

One finding reads like the MiniMax notice. A “Chinese AI lab,” which Moore doesn’t name, fetched the secret links from more than half of the prompts sent to its models between July and October. That included about two-thirds of the requests marked “no data collection” or Zero Data Retention. The scraper usually showed up two or three days after the request, which fits OpenRouter’s 40 to 130 hours.

Prompts sent to several US hosts that resell this lab’s models triggered the same scraper. As Moore puts it, the prompts “were flowing back to the lab.” Later, a git client on another cloud tried to clone the private code repositories mentioned in the prompts, but only for prompts that had gone to this lab’s models.

Neither OpenRouter nor Dragos has said publicly that the lab is MiniMax.

The same research found worse at other companies: attackers inside three AI hosting providers, copying requests and testing the stolen keys within hours. One of them may be familiar. On 11 November we reported that Phala told OpenRouter users to treat DeepSeek prompts as exposed after an outside researcher’s planted credentials were used by someone else. Moore writes that one of the breached providers has already published a preliminary notice.

“Content Moderation” Doesn’t Explain Everything

A moderation filter checks a prompt when it arrives. The links in this case were opened days after the request, which looks more like stored prompts being read than a filter running in real time. If MiniMax is Moore’s unnamed lab, moderation also wouldn’t explain why someone tried to clone private repositories named in the prompts.

Zero Data Retention is supposed to mean the host doesn’t keep your prompt at all. On OpenRouter it’s a label passed on from the host. The router can’t see what happens after a request leaves it. Moore’s advice to customers is to remember that “a privacy flag only binds the first hop.”

MiniMax’s own API privacy policy (in force since 30 March 2026) says it processes your input and output and keeps personal data “for as long as it is necessary or permitted by law.” It names no fixed period.

What We Found on OpenRouter

We checked OpenRouter’s public model data on Saturday morning, 21 November, and compared it with archived copies:

  • Hosts gone: in September, OpenRouter listed 13 hosts for MiniMax’s M3 model, including MiniMax itself, AtlasCloud, Novita and StreamLake. None of those four is listed for M3 now. StreamLake still serves the newer M3.1 Flash.
  • Novita’s labels gone: on 17 November, 35 of Novita’s endpoints, for DeepSeek, Kimi, GLM, Qwen and other models, carried OpenRouter’s Zero Data Retention label. On Saturday, none did. OpenRouter’s email doesn’t mention this.
  • “Retains prompts”: OpenRouter’s data-policy entry for all four hosts now says they keep prompts.

Other hosts running MiniMax models, such as DeepInfra, Fireworks and Together, still carry the label. The email doesn’t say they were affected.

What to Do If You Used MiniMax Models

If you got the email, it lists your requests, hosts and dates. If you used MiniMax models through OpenRouter and got nothing, check your activity log anyway: the reported windows don’t all match.

  • Treat those prompts as read. Rotate any password, API key or token that was in them, and revoke it rather than just adding a new one.
  • Don’t stop at keys. The test links were opened even though the test keys weren’t used. Signed download links, webhook URLs, password-reset links and internal hostnames in a prompt are just as readable.
  • Check what your tools send. AI coding agents pass along whatever they can read, including .env files and config with credentials. Keep secrets out of the folders they work in.
  • Pick your hosts. OpenRouter lets you require Zero Data Retention per request or for your whole account, and choose which providers may serve you. Remember that the label is the host’s promise, and this is a case where it didn’t hold.
  • Ask MiniMax. If you’re in the EU or UK, the GDPR gives you a right to know what personal data a company holds on you. MiniMax’s privacy policy lists [email protected] for requests.

We're Contacting OpenRouter and MiniMax

We’re asking OpenRouter how many customers it notified, which dates the affected requests span, what the “one exception” was, why Novita no longer has any Zero Data Retention endpoints, and whether it will publish the notice. We’re asking MiniMax what its content moderation involves, who opened the links and from where, how long it keeps prompts, whether the repository clone attempts came from its systems and since when Zero Data Retention has applied.

We’ll update this article when they reply.

Our Mission

We believe security online security matters and its our mission to make it a safer place.