Distillation attack
Named by 5 kept items in this workspace.
3 items name it under classification.key entities and 2 under techniques.
The items that name it 5
Covert, industrial-scale extraction of one model's capabilities to replicate them in another without authorization, often using networks of fraudulent accounts
Attackers systematically collect large volumes of frontier model reasoning traces and train (distill) their own models on these traces to replicate state-of-the-art performance at low cost
In every workspace you can read 8
Security
All techniques in this workspace.