OpenAI disrupts Moonshot-linked reasoning extraction campaign

OpenAI said it had identified and disrupted a coordinated distillation campaign aimed at extracting protected reasoning from its models. It said a core cluster of the activity went back to the first week of July and was associated with people linked to Moonshot AI, but it did not cite technical evidence for that judgement.

The company said the activity began on 1 July 2026 and remained low at first before rising sharply on 24 and 25 July. OpenAI said it saw 16,000 attempted requests from more than 4,000 users and later found related prompt-pattern activity across more than 15,000 users.

OpenAI said it fully disrupted the campaign on 28 July 2026. It said it had since added mitigations, banned the accounts involved and closed a pathway that could let someone replay encrypted reasoning if they already had it.

The company said protected reasoning could expose sensitive data and help others reproduce model capabilities. It also said adversarial distillation could create safety and national security risks by transferring advanced capabilities without the safeguards applied to user-facing outputs.

In August 2026, researchers from MATS Research, ELLIS Institute Tübingen and Synk published a study on an architectural vulnerability affecting Claude, Gemini and GPT. Anthropic also accused Moonshot AI last month of passing customer requests to Claude rather than processing them with Kimi, and said some exchanges were retained for training.