Most AI agent guardrails guard against danger. Almost none guard against an agent sounding like someone else. Here's a framework for the ones that do.