Models that introspect on their own thoughts could be a huge boon for researchers - or pose a threat to humans.