OpenAI reported a campaign aimed at extracting protected reasoning from its AI models, which began on July 1. The activity peaked between July 24 and 25, involving approximately 16,000 requests from over 4,000 users. OpenAI blocked the campaign by July 28.
The company identified the activity as unauthorized distillation, where operators manipulated interactions to force models to reveal hidden reasoning. OpenAI linked the core of these actions to individuals associated with the Chinese developer Moonshot AI, though it noted that not all operators necessarily belonged to the same entity.
In response, OpenAI restricted affected accounts, enhanced registration checks, and improved infrastructure monitoring. The company also implemented technical barriers to prevent the reuse of encrypted reasoning blocks across different conversations and strengthened isolation between user workspaces.
OpenAI shared its findings with other AI firms via the Frontier Model Forum and government channels. This follows previous allegations from Anthropic, which claimed in February that Moonshot AI and other developers conducted large-scale extraction campaigns against its Claude models.
