Prompts sent to MiniMax through OpenRouter were opened 40 to 130 hours after delivery, against a promise of zero retention, per StartupHub.ai reporting. The test used single use canary URLs planted in enterprise prompts, and clicks came days later from third parties with no local access needed. The affected path covered MiniMax endpoints reached via OpenRouter between July and , with OpenRouter saying its own systems were not breached.
Zero retention meant five days of storage.
A researcher laid the trap with unique URLs that should have died on delivery. Instead third parties fetched them 40 to 130 hours later, which exposed logging far outside the API call. MiniMax then told intermediary hosts including StreamLake that content moderation was enabled before , so prompts passed through brokers such as AtlasCloud and Novita into a review pipeline. That explanation fits Chinese compliance, where the Cyberspace Administration of China requires generative AI providers to monitor and block illegal information. The result is that a OpenAI model in the US and a MiniMax model in Shanghai looked interchangeable in one dropdown, but carried opposite retention law.
Phala, DeepSeek and the 18 day detour
MiniMax is not the only OpenRouter route that leaked. Phala published a preliminary incident report saying an attacker held a server running DeepSeek V4 Flash 0731 for 18 days while OpenRouter sent it chatbot prompts. That window included passwords, keys and other prompt text, a remote exposure that needed no victim click. Put next to 40 to 130 hours of moderation retention at MiniMax, the read is blunt. One was state required review, the other was criminal takeover, but both flowed through the same abstraction.
A second comparison comes from Ars Technica reporting on a 2m canary experiment across widely used AI services, which found sensitive data leaking like a sieve. The scale matters because it shows MiniMax was not an edge case. The performance story hid the risk for a time, with MiniMax Flash praised for coding wins days before the retention notice. Benchmarks measured speed and score while jurisdiction decided who could read the prompt.
The numbers do not show how many prompts were read by a human, how many were scanned by machine, or how many contained live AWS URLs, database tokens or PII. OpenRouter has not published a customer count and MiniMax has not detailed its reviewers. A full per customer log of what was retained and who fetched it would change the read from assume breach to scoped breach. Until then startups should rotate any credential sent to MiniMax in that July to window and scrub PII before it leaves their own servers.
How many production keys sent in July are still valid after being readable for 40 to 130 hours with no alert?





