OpenAI accuses Moonshot AI associates of attempting to extract GPT reasoning

OpenAI has announced the blocking of a coordinated campaign aimed at extracting the protected internal reasoning processes of its GPT models. The company traced the core group of actors to individuals associated with Moonshot AI, a Chinese startup known for developing the Kimi chatbot.

The campaign, which began in early July, saw a significant surge on July 24 and 25, with 16,000 extraction requests from over 4,000 users. In total, the activity involved more than 15,000 users before OpenAI successfully blocked the operation on July 28.

According to OpenAI, the attackers did not breach encryption or access databases. Instead, they manipulated model interactions to replicate protected reasoning—the internal step-by-step logic models use before providing a final answer. One method involved copying encrypted reasoning from one conversation and requesting the model to decode it in a separate session. OpenAI described this unauthorized practice as “adversarial distillation,” a technique used to train smaller models using the outputs of more powerful ones.

OpenAI has implemented several countermeasures, including patching the mechanism that allowed the replay of encrypted reasoning and restricting accounts involved in fraudulent activity. While OpenAI has named Moonshot AI, the company noted it has not provided specific technical evidence for the attribution due to security reasons.