OpenAI Accuses Moonshot AI of Organized Adversarial Model Distillation Campaign
OpenAI announced that it detected and blocked a coordinated adversarial model distillation campaign on its platform involving over 15,000 user accounts designed to extract model reasoning and outputs. The company publicly linked the core activity group behind the campaign to Moonshot AI, the Chinese AI startup behind the Kimi model. This public accusation highlights mounting friction over intellectual property, API terms of service enforcement, and competitive model scraping between top US and Chinese AI labs. It marks a key moment in AI governance, demonstrating how leading frontier AI labs use joint threat-sharing mechanisms to disrupt unpermitted model distillation. The distillation activity began in early July and surged around July 24–25 when over 4,000 accounts made 16,000 requests, leading to a complete block by July 28. OpenAI emphasized that no internal databases or encryption were breached, and it shared technical indicators with industry peers via the Frontier Model Forum to strengthen collective defenses.
## BACKGROUND
Model distillation is a technique where a smaller, more efficient 'student' model is trained using outputs generated by a larger, more capable 'teacher' model like GPT-4. Adversarial distillation occurs when this extraction is conducted systematically without the provider's permission, often violating API usage policies. The Frontier Model Forum is an industry organization founded by frontier AI developers to facilitate information sharing on safety risks and security threats.