OpenAI’s misalignment monitoring system had “a number of novel vulnerabilities,” third-party evaluators found.