Model extraction reconstructs or approximates a private model through repeated API queries. An attacker studies the responses and trains a student model to mimic the target’s behavior; the result need not reproduce its weights exactly to expose its functional logic. section 4.4.3
Unlike direct theft, which accesses stored model weights, extraction targets the model’s behavior through its interface. This distinction matters because a secured storage environment does not by itself prevent functional imitation. section 4.4.3
Reasoning-model imitation
Sky-T1-32B-Preview was reported to perform on par with o1-preview on popular reasoning and coding benchmarks using under $450 in compute. OpenAI and Microsoft also investigated whether DeepSeek used large volumes of OpenAI API data to train competing models. section 4.4.3
