How are you sandboxing your agents? You ARE sandboxing your agents, right? Right?!
Put another way: What is the worst-case scenario if an attacker compromises an NPM package and your agent downloads and runs it?
Model alignment and classifiers will not save you.