A survey-focused look at enterprise AI agents points to four critical problem areas, and the broader technical lesson is clear: once software can act, trust boundaries become the real attack surface.