Large language models (LLMs) are increasingly integrated into healthcare, reshaping the organization and ethical norms of traditional mental health counseling. General-purpose LLMs lower access barriers through broad linguistic capabilities, while domain-specific LLMs aim to improve task adaptability through specialized training for mental health scenarios. This emerging paradigm shift, featured by data-driven approaches, algorithm-enabled mediation, and scalable service delivery, may expand availability, reduce cost, and support proactive prevention. However, such technological integration also introduces ethical and safety risks, including privacy breaches in data collection and retraining, hallucinated or misleading outputs, and technological paternalism that may erode users' informed consent and autonomy. In addition, amplified biases and over-reliance may cause subtle harms and severe clinical consequences in extreme cases. Therefore, governance should adopt a differentiated approach and regulate key stages of model development, deployment, and service delivery across four dimensions, including strengthening data security, improving algorithmic governance guidelines, reaffirming human-centered values, and enhancing professional ethical standards, to enable safe, controllable, and equitable applications of LLMs in mental health counseling.