Maybe this is an elementary angle given my lack of security experience, but couldn't Microsoft figure out a wat to parse the documents prior to model analysis/action? Implement some form of deterministic layer that resides between the user and the model?