“During over 3,000 estimated hours of red teaming, no universal jailbreak was found that could extract information from an early Constitutional Classifier-guarded LLM at a similar level of detail to an unguarded model.”