OpenAI Disrupts AI Data Extraction Campaign Linked to Moonshot AI
OpenAI has announced that it blocked a campaign that was unauthorizedly extracting protected reasoning data from its AI models. The activity, which took place in July 2026, was determined to be conducted by individuals linked to Beijing-based Moonshot AI.
OpenAI announced that it has detected and stopped a coordinated 'adversarial distillation' campaign targeting the operational logic and protected reasoning processes of its AI models.
The campaign, which began on July 1, 2026, and was completely terminated on July 28, aimed to manipulate model interactions to reproduce protected reasoning processes in a visible format. OpenAI stated that the individuals behind these activities were linked to the Beijing-based company Moonshot AI.
The campaign peaked on July 24 and 25 with 16,000 attempts carried out across more than 4,000 users. Additionally, prompts with similar patterns were detected in over 15,000 users. OpenAI emphasized that there was no database breach or encryption compromise, but that terms of service were violated.
The company took measures by closing fake accounts and blocking the path that allowed for the replay of encrypted reasoning traces. OpenAI warned that such methods pose risks to national security and model safety, as they could lead to the training of models without original security safeguards.
Sources
- thehackernews.com · Link