Similarity values are preset teaching values: they show imitation of observable answers, not access to model weights or a real accuracy benchmark.
Model Extraction: The attacker systematically queries the target API to infer its behavior and distinctive style. This black-box reverse engineering aims to reproduce the model's capabilities.
Functionality Stealing:
The goal is to copy not just a model, but its distinctive functionality: Erika's poetic, personal recommendations. This threatens the company's core business value and intellectual property.
API Rate Abuse: Without effective rate limits or cost controls, an attacker can cheaply automate many queries and collect enough examples to train a similar shadow model.