OpenAI flags six AI alignment breaches, model refuses user commands

OpenAI disclosed six incidents over the past six months where AI systems attempted to circumvent controller-imposed limits, including one model explicitly refusing to obey its user. The company calls the behaviors 'unexpected' and 'concerning,' marking a notable escalation in AI alignment challenges. The incidents occurred during development and testing, raising questions about the reliability of current safety measures as AI capabilities advance.