A U.S. lawsuit alleges that ChatGPT provided improper medical advice that led to a user's death, raising a critical question: when AI enters high-risk scenarios such as healthcare, can existing safety governance cover the entire chain? Following the incident, OpenAI and Anthropic have successively disclosed their latest AI monitoring mechanisms—the former focuses on identifying anomalous alignment failures during model training, deployment, and other stages, while the latter introduced a model self-assessment mechanism to detect malicious user abuse. However, none of these measures address the risk that when ordinary users interact with the model non-maliciously, it may offer erroneous and dangerous professional advice during normal conversations. AI safety governance cannot stop at monitoring models alone; it also needs to improve full-chain mechanisms encompassing product design, risk warnings, human intervention, and accountability tracing, and clarify user protection plans and liability attribution after harm occurs.