ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI Q&A
OpenRouter Warning: Prompts Sent to MiniMax May Have Been Read by a Third Party

OpenRouter Warning: Prompts Sent to MiniMax May Have Been Read by a Third Party

AI Q&A • Admin • • 8 views

OpenRouter warning: prompts sent to MiniMax models may have been read by a third party. Note the word "may" — this is not a confirmed theft — but OpenRouter's advice is direct: treat any secret that appeared in those prompts as already leaked. On November 21, 2026, OpenRouter emailed affected users with the subject line "Requests you sent to MiniMax models may have been read". The problem is not that OpenRouter was hacked, but the MiniMax side of the chain.

How it was discovered

Between July and October 2026, researchers sent about 2 million bait prompts to several models, each containing a unique test link or fake credential. For prompts sent to one unnamed "Chinese AI lab", more than half of the secret links were opened by a third party 40–130 hours later, and about two-thirds of those had been marked zero data retention. OpenRouter's email named that lab as MiniMax — the original research did not name it, and the link between the two comes from cross-checking by several media outlets. Whether MiniMax served the request directly or through AtlasCloud, Novita or StreamLake, the result was the same; prompts sent to other models in the same batch were essentially never accessed.

MiniMax's explanation, relayed through StreamLake, is that content moderation was enabled before October 22 and zero data retention has been enabled since. OpenRouter says it has asked MiniMax for its own statement and had not received one at the time of writing, and it has already removed the zero-data-retention label from the affected MiniMax endpoints.

Three things you should do

First, check your OpenRouter activity log for every call sent to MiniMax models since July, including those routed through AtlasCloud, Novita and StreamLake. Second, rotate or revoke any passwords, API keys and tokens that appeared in those prompts — the researchers' AWS test key was never used, which confirms prompts were stored and read later rather than credentials being stolen, but treating them as leaked is still the safest course. Third, stop sending sensitive content to MiniMax models.

One more reminder: do not treat any platform's zero-data-retention label as a compliance guarantee. OpenRouter's own data-processing agreement states that zero data retention only reflects the provider's self-declared policy. Truly sensitive content should never have been in a prompt sent to a third-party model in the first place.

Recommended Tools

More