Frontier AI models utilize automated post-filtering that cuts off user interactions when prompts trigger sensitive cybersecurity or biological safety boundaries