Anthropic has announced the identification of large-scale 'distillation' attacks targeting its language models. The attacks were reportedly conducted by DeepSeek, Moonshot AI, and MiniMax.

The 'distillation' technique involves training a smaller model (the 'student') using the output of a larger, more complex model (the 'teacher'). The goal is to transfer the capabilities of the larger model to a lighter one, resulting in a model that is more efficient in terms of computational resources.

This type of attack raises significant concerns about the intellectual property and security of AI models. If a model can be effectively 'distilled', it becomes easier for third parties to replicate its functionality without having to train it from scratch.