OpenAI disrupts novel AI distillation attack it links to China

OpenAI announced it has stopped a coordinated attempt to extract and distill reasoning capabilities from its AI systems, attributing a main part of the operation to associates of China-based Moonshot AI. Attackers exploited a method where encrypted data from one conversation could be decrypted in another, exposing protected model outputs. OpenAI has blocked the accounts involved, tightened sign-up controls, and fixed the flaw. The company warned similar vulnerabilities may exist in other AI models.

• Attackers extracted reasoning data via a cross-session decryption method
• OpenAI linked activity to Moonshot AI, a Chinese rival
• Over 15,000 suspicious user accounts identified and banned
• OpenAI improved monitoring and infrastructure controls
• Vulnerability may persist in other AI models, OpenAI warned peers