AI agent guardrails are a necessary starting point for enterprise AI safety — but they have known bypass paths, including prompt injection, multi-agent permission creep, and silent drift after model updates. This post maps the failure modes and explains why runtime governance is the layer most enterprise teams are missing.