OpenAI blocks large-scale inference extraction operation

On September 30, OpenAI disclosed that it had detected and disrupted a large-scale "adversarial distillation" operation. The activity first appeared as early as July 1, with operators systematically invoking the model in an attempt to extract protected reasoning that would not normally appear in responses, for use in training, replicating, or improving other models. Between July 24 and 25, OpenAI monitored approximately 16,000 related requests from over 4,000 users; further investigation revealed that similar prompt patterns involved more than 15,000 users, and the company completed its response by July 28. The attack method included copying encrypted reasoning from one conversation and then asking the model in another conversation to decrypt and transcribe its content. OpenAI emphasized that the operators did not crack encryption, breach databases, or directly access stored user conversations. The company cannot yet confirm whether all activity originated from the same actor, but it attributed one core cluster to individuals associated with Moonshot AI, the developer of Kimi. OpenAI has banned or restricted the relevant accounts, strengthened registration, infrastructure, and streaming output detection, and patched pathways for replaying encrypted reasoning across users, workspaces, and models. Relevant information has also been shared with partners through the Frontier Model Forum and government information-sharing channels.