A stress test of ChatGPT Health triage revealed missed high-risk emergencies and inconsistent activation of suicide-crisis safeguards, raising safety concerns for consumer-scale deployment.