OpenRouter Warns MiniMax Prompts May Have Been Read
Email says a third party may have accessed prompts and outputs sent to MiniMax models through OpenRouter, which has removed a data-retention label while it investigates.
Updated on Nov. 21, 2026 in Cybersecurity

Live Poll
Do you trust AI services to keep what you type private?
OpenRouter emailed users warning that prompts and outputs sent to MiniMax models may have been read by a third party. A security researcher's canary-link test found third-party fetches 40 to 130 hours after requests were sent, and OpenRouter says its own systems were not breached.
Why it matters
The findings suggest sensitive prompts sent through third-party forwarding providers may be retained and accessed beyond what users expect. MiniMax has tied prompt retention to content moderation compliance with China's Cyberspace Administration, but neither company has published a full log of what was retained or who accessed it.
Canary links placed in test prompts between July and October 2026 were fetched by a third party 40 to 130 hours later, a pattern that held whether MiniMax served the request or a forwarding provider did. The full scope of retained data and the identity of the accessing party remain unknown.
The players
OpenRouter
A platform that routes API requests from developers to a range of AI model providers.
MiniMax
A Chinese artificial intelligence company that develops large language models.
The details
The OpenRouter email names AtlasCloud, Novita and StreamLake as forwarding providers. MiniMax told StreamLake that content moderation was enabled on its endpoints before October 22, and OpenRouter has removed the Zero Data Retention label from affected endpoints while it investigates.
Timeline
July 2026: Canary-link test prompts to MiniMax begin.
October 2026: Canary-link test prompts end.
October 15 to November 12, 2026: Affected period reported by RuntimeWire.
October 22, 2026: Content moderation reportedly enabled on MiniMax endpoints before this date.
November 21, 2026: OpenRouter sends warning email to users.
Under the Hood
The retention practices described are tied by MiniMax to content moderation compliance with China's Cyberspace Administration, linking a national AI regulatory regime to data handling by intermediary platforms. The episode shows how regulatory obligations in one jurisdiction can shape what happens to prompts routed through a global marketplace.
Developers and users who sent prompts through OpenRouter to MiniMax models may have had that content retained and accessed by third parties. Anyone relying on the Zero Data Retention label should check whether affected endpoints are in use until OpenRouter completes its investigation.
The takeaway
Users sending sensitive material through intermediary AI platforms should confirm which providers actually handle their data, since a retention label may not cover every hop. Until full logs are published, treating third-party routed prompts as potentially retained is the cautious course.
Further reading
Follow related coverage in Cybersecurity for more on how data handling risks play out across AI platforms.
Source note: This article includes information reported by Startup Fortune.
Live Poll
Do you trust AI services to keep what you type private?



