OpenAI has disrupted a coordinated campaign by Beijing-based Moonshot AI that attempted to illicitly extract protected reasoning from its artificial intelligence models through adversarial distillation techniques.
Key Points
- OpenAI identified and blocked a campaign that used over 15,000 user accounts to extract model reasoning between July 1 and July 28, 2026.
- The activity involved manipulating model interactions to reproduce protected reasoning, violating OpenAI’s terms of service.
- Researchers from MATS Research and the ELLIS Institute identified an architectural vulnerability allowing encrypted reasoning traces to be decrypted across different sessions.
- OpenAI implemented new mitigations and closed security pathways to prevent the replay of encrypted reasoning data.
- Moonshot AI previously faced similar accusations from Anthropic regarding the unauthorized use of Claude’s outputs to train its own models.