Anthropic just raised the alarm: it says a network tied to Chinese AI companies created thousands of accounts to hammer Claude with prompts, capture the outputs, and use them to train competing models—a technique known as model distillation. If true, this is a new kind of "theft": not a server hack, but a high-volume scraping of a model's capabilities through the front door of an API. In this video, we break down what distillation is, why it threatens frontier-model IP, and how it can quietly shift the balance in the global AI race. We'll cover the alleged tactics (fake accounts, automation, evading rate limits), the defenses (behavioral detection, throttling, watermarking, and policy changes), and the bigger stakes for startups, national security, and open research. Most importantly, we'll ask: how do you protect an AI model when the product is an interface and the "asset" is knowledge embedded in weights? Watch to the end for practical safeguards any team shipping an LLM can implement today—and what this episode signals about what comes next. We'll also separate hype from evidence, explain what companies can log without violating privacy, and discuss whether regulation, contracts, or technical friction will matter most in the future.