OpenAI ties a reasoning extraction campaign to Moonshot AI

The activity began in the first week of July. It peaked on 24 and 25 July with 16,000 requests from more than 4,000 users, and by 28 July a cluster of over 15,000 users was sending similar prompts, Wccftech reports.

The method was new. Operators copied the encrypted reasoning from one chat and asked the model in another chat to decode it. OpenAI says the encryption itself held and no database was breached. Its fixes are account bans, tighter sign-up checks, more monitoring and a change that stops one user’s encrypted reasoning from being read in another user’s session, BankInfoSecurity adds.

The practice, called distillation, trains one model on another’s outputs. Anthropic has made similar claims against Moonshot, BankInfoSecurity notes, and the published reports carry no reply from Moonshot.

For buyers of cheaper open models, provenance is now part of the risk: two US labs say people tied to one well-known rival harvested their models’ reasoning.