An advanced artificial intelligence model developed by Anthropic identified vulnerabilities in highly sensitive U.S. government computer systems within hours during an authorized security test, according to officials.