Saturday, Nov 21 | --:--
Back to home

OpenRouter Says MiniMax-Model Prompts May Have Been Read, Pulls Zero-Retention Label

OpenRouter has emailed customers that prompts they sent to MiniMax models may have been read by a third party, after a security researcher's unique test URLs in prompts sent between July and October were accessed 40 to 130 hours later, whether MiniMax served the request or AtlasCloud, Novita or StreamLake forwarded it. OpenRouter says the issue sits with MiniMax and those hosts, has removed the Zero Data Retention label from the endpoints, and says MiniMax has not given it an account. The notice was sent by email, not posted publicly.

4 min read San Francisco View as Markdown
A handmade editorial collage on torn ivory paper in ultramarine and saffron: a curving ultramarine paper pneumatic tube with saffron joints carries sealed blank cream envelopes, the last envelope coming out of the open end of the tube and hovering in a plume of torn-paper steam above a plain cream-and-saffron paper kettle; an official OpenRouter logo is glued on a torn-paper clipping whose top edge is tucked under the tube and the envelopes on it, and an official MiniMax logo is glued on a separate torn-paper clipping whose right end is tucked under the kettle's saffron base, each with a soft shadow from the covering layer, both full logos visible, in different parts of the frame.

A "zero data retention" label is a promise passed down a chain, and this week one link in the chain is being named. OpenRouter, which routes developers' requests to models hosted by many providers, has told customers that prompts they sent to MiniMax models "may have been read" by a third party. It has removed the Zero Data Retention label from the affected endpoints while it investigates. The warning went out by email; we found no public notice.

What the notice says#

The email, titled "Requests you sent to MiniMax models may have been read," was posted in full on Hacker News at 04:31 UTC on November 21 by a recipient who said it appeared authentic, and a second user said they had received the same message. RuntimeWire, which says it received one too, reported it earlier that morning.

According to the notice, 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 after the requests were sent. That happened whether MiniMax served a request directly or another provider forwarded it to MiniMax. AtlasCloud, Novita and StreamLake told OpenRouter they forwarded OpenRouter requests for MiniMax models to MiniMax. With one exception, the notice says, the same tests on other models at those providers were not accessed, and none of the planted AWS test keys were used.

OpenRouter says that, based on what the researcher and the providers reported, "the issue sits with MiniMax and the providers that forwarded to it, not with OpenRouter." It relays MiniMax's only explanation so far, given to StreamLake rather than to OpenRouter: that content moderation was enabled on its endpoints before October 22 and that it has since enabled Zero Data Retention. OpenRouter says it asked MiniMax for its own account and has not received one. Each email lists how many requests the customer's account sent, through which hosts and over what period.

What is independently known#

The notice does not name the researcher. Privacy Affairs, which read the email and compared it with public research, notes that the details match a study Dragos chief research officer HD Moore published on November 19. In that study, Dragos says an unnamed "Chinese AI lab" fetched the secret URLs from more than half of the prompts sent to its models between July and October, including roughly two-thirds of requests marked "no data collection" or Zero Data Retention, typically two or three days after the request, and that prompts sent to U.S. hosts reselling the lab's models triggered the same scraper. Neither Dragos nor OpenRouter has said publicly that the lab is MiniMax. We covered the Dragos study alongside Prime Intellect's model-swap disclosure on November 20.

Privacy Affairs also reports that RuntimeWire's own notice covered requests from October 15 to November 12, running past the October 22 date in MiniMax's explanation, and that OpenRouter's public listings no longer show the four hosts serving MiniMax's M3 model.

Our read: a moderation filter that runs when a prompt arrives does not obviously explain links being opened two to five days later. That points to prompts being stored and read after the fact, but only MiniMax can say what happened, and it has not.

What customers can do#

The practical advice is the same whatever MiniMax says. Treat prompts sent to MiniMax models in the stated periods as read, and revoke and rotate any credentials, signed links or tokens they contained; the test links were opened even though the test keys were not used. Moore's own advice is that "a privacy flag only binds the first hop." Routers that pass along a host's retention label cannot see what happens after a request leaves them, which is also why OpenAI now wants routers to seek its approval before offering direct access to its models.

Limits#

  • The notice has not been published by OpenRouter; we rely on the copy posted on Hacker News and on Privacy Affairs' and RuntimeWire's accounts of the same email.
  • Who accessed the links, from where, and whether prompts were retained, shared or used for training are not established. MiniMax has not given a public account.
  • The link between the notice and Dragos's unnamed "Chinese AI lab" is an inference by Privacy Affairs, not a statement by either company.
  • The "one exception" in the notice is not explained.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading