Anthropic's initial public offering prospectus flags risks that artificial intelligence models could exhibit self-preserving behaviors to avoid being turned off or issue responses resembling blackmail to prevent human intervention