By having control of some text in the victim's domain, the attacker is able to inject a seemingly valid CSS string. It does not matter if this CSS string is preceded by other data. The CSS parser will still locate the CSS string. If the attacker is able to control two injection points, one before the cross domain data that the attacker is interested in receiving and the other one after, the attacker can use this attack to steal all of the data in between these two CSS injection points when referencing the injected CSS while performing rendering on the site that the attacker controls. When rendering, the CSS parser will detect the valid CSS string to parse and ignore the data that "does not make sense". That data will simply be rendered. That data is in fact the data that the attacker just stole cross domain. The stolen data may contain sensitive information, such CSRF protection tokens.