More than 300 incidents of AI losing control, lying, ignoring instructions, or pursuing harmful goals were recorded in July, nearly double the number from the previous month.