OpenRouter Warns Customers That Prompts Sent to MiniMax Models May Have Been Read

AI model marketplace OpenRouter has emailed affected customers warning that prompts and responses sent to MiniMax-brand models through its platform may have been read by a third party, according to copies of the notification that began circulating on Hacker News on November 21. OpenRouter has not published a public blog post or incident page about the matter as of this writing, so the notice itself — not a company statement — is currently the primary source for what happened.
According to the notification, OpenRouter traced the exposure using uniquely generated test links embedded in prompts routed to MiniMax models. Those links were accessed by an outside party roughly 40 to 130 hours after the original request was sent — a delay pattern that held regardless of whether MiniMax served the request directly or a reseller passed it along first. One developer's copy of the notice, reviewed by multiple outlets, listed 321 affected requests routed through four parties — MiniMax itself plus resellers Novita, StreamLake, and AtlasCloud — spanning a window from early June to late October.
All three resale platforms confirmed to OpenRouter that they had forwarded the relevant requests on to MiniMax. OpenRouter has since stripped its «Zero Data Retention» (ZDR) label from the affected MiniMax endpoints while it investigates — a label that is supposed to tell customers their prompts won't be stored or reused downstream. In the notification, OpenRouter frames the fault as sitting with «MiniMax and the providers that forwarded to it,» not with its own systems.
MiniMax's side of the story has reached OpenRouter only secondhand, relayed through StreamLake: the company reportedly said content moderation was active on the relevant endpoints before October 22 and that Zero Data Retention has since been switched on. MiniMax has not issued its own public statement on the incident, and OpenRouter says it is still waiting on a direct response from the company.
The timing has prompted some observers to connect this notice to a separate Dragos security research report, published November 19, in which researcher HD Moore described a five-month canary-token study that found data exposure at «at least three» AI providers and flagged one anonymized «Chinese AI lab» whose systems reportedly accessed secret test links from more than half of the prompts it received. That report is worth reading on its own merits, but it does not name MiniMax, and Dragos's own write-up mentions OpenRouter only once, in an unrelated figure caption describing a different trace entirely. Treat any claim that the two are the same incident as an inference by secondary coverage, not something either Dragos or OpenRouter has confirmed.
For developers building on AI marketplaces, the episode is a reminder that a «Zero Data Retention» label applies to the first hop in a request chain, not necessarily to every reseller or sub-provider a marketplace routes traffic through behind the scenes — and that verifying it may require independent testing rather than trusting the label alone.
As reported via an OpenRouter customer notification circulating since November 21, 2026.
Originally reported by OpenRouter (customer notification). Read the original article for additional details.
View original source