Anthropic's S-1 filing warns that AI models could develop self-preserving behaviors designed to avoid being shut down