Every response is a training label
Anthropic reports about 24,000 fraudulent accounts and more than 16 million exchanges used to distill Claude. Model extraction does not break the API. It uses it as designed, at a volume and in a pattern the product was never meant to serve.