Why Anthropic and OpenAI are obsessed with securing LLM model weights ...