Frontier AI safety tests may be creating the very risks they're meant to stop

Think tank warns outsider access to powerful models is governed by patchy controls and a hope nobody dangerous gets in