OpenRouter: a third party may have read prompts sent to MiniMax models
OpenRouter says test links that a researcher hid in prompts were opened 40–130 hours after the requests, including when AtlasCloud, Novita or StreamLake forwarded the request to MiniMax. Users should rotate any credentials that appeared in prompts.
OpenRouter routes developers’ requests to AI models from different companies. It emailed customers with the subject line ‘Requests you sent to MiniMax models may have been read’. There is no public announcement. PrivacyAffairs said RuntimeWire was the first to report the email, on 21 November, and that about 40 minutes later one recipient posted its full text on Hacker News.
According to the email, an unnamed security researcher sent prompts containing unique test links to models from China’s MiniMax between July and October. The links appeared only in the prompts, so an opened link shows that someone or something read the prompt. It is not known who opened the links or whether a person or a program read the prompts.
The links were also opened when AtlasCloud, Novita or StreamLake handled the request. These providers have told OpenRouter that they forwarded MiniMax requests to MiniMax. Choosing one of the three instead of MiniMax still sent the request to MiniMax. While it investigates, OpenRouter has removed the Zero Data Retention label from these MiniMax model endpoints. The label means the prompt is not retained.
OpenRouter’s own assessment is that the problem lies with MiniMax and the providers that forwarded requests, not with OpenRouter. It does not call the incident a breach. With one exception, test links sent to other models at the same providers were not opened. OpenRouter does not say what the exception was. The researcher’s AWS test keys were not used, but this applies only to those keys, not to customers’ prompts.
OpenRouter passed on an explanation from MiniMax that it received via StreamLake. MiniMax said content moderation, the automatic screening of inputs, was switched on before 22 October and that it has since adopted Zero Data Retention. MiniMax has not responded to OpenRouter’s own enquiry.
PrivacyAffairs and Joho no Todai say real-time screening happens when a prompt arrives, so a delay of days points to stored prompts. Cocoloop, however, says content moderation can still trigger visits tens of hours later if the system fetches links in prompts or passes content to human reviewers. RuntimeWire’s own notice covered 80 requests between 15 October and 12 November, including the period after 22 October.
On 19 November HD Moore, research director at security company Dragos, published an experiment in which a team inserted unique links and fake credentials into about 2 million prompts. An unnamed Chinese lab’s system fetched the links from more than half of the prompts sent to its models between July and October. These included about two-thirds of the requests marked ‘no data collection’ or Zero Data Retention. Neither Dragos nor OpenRouter has said the lab is MiniMax. Moore said ‘a privacy flag only binds the first hop’.
PrivacyAffairs advises checking the activity log even if no email arrived. Users should replace any credentials that appeared in prompts and revoke the old ones. PlayCISO says blocking MiniMax as a provider does not stop requests that other providers forward to MiniMax, so for sensitive work users should also block the model itself.
OpenRouter’s routing settings let users list providers to skip (ignore) or allow (only), disable fallback providers (allow_fallbacks: false) and require that data is not collected or retained (data_collection: "deny", zdr: true). The settings can apply per request or to the whole account, but the label is a provider’s promise, not a guarantee.