Comment by jdthedisciple

simple solution:

preprocess any input to agents by restricting them to a set of visible characters / filtering out suspicious ones