Mitigating LangSec Problems With Capabilities