Abliterated Models in Production: What Removing Safety Guardrails Means for Agent Security Boundaries