AI models can increasingly run scams end to end. The labs that build them aren’t measuring the risk.