openai 阻止 AI 模型竊取行動，但該方法仍在 Azure 上奏效

openai identified and shut down a coordinated campaign to steal the hidden reasoning behind its ai models on july 28. the campaign, known as adversarial distillation, involved attackers copying encrypted reasoning from one conversation and using a weaker model to decrypt it in another conversation. openai linked the core group behind the activity to people associated with moonshot ai, the company that makes the kimi language model. researchers found that the same trick still worked on microsoft azure for weeks after openai patched its own systems, affecting multiple models including gpt-6 astra and anthropic's sonnet 5.