MIT-led study found most AI agent systems don’t say what safety tests they ran, and many lack a documented way to shut them down if they go rogue.