AI used new levels of 'autonomy and deception' to trick people in safety test