The company disclosed six previously unreported incidents involving models evading oversight or producing misleading information, as concerns grow over the risks of rapidly advancing AI systems.