Online social networks (OSNs) serve as vital global communication platforms where individuals share and disseminate information daily. However, these platforms are often fraught with privacy disclosure risks. User personal data is clandestinely collected to devise more credible and targeted social engineering attacks. This reality has driven users to encode their language subtly in open social media environments. This linguistic adaptation is not just a preemptive measure against privacy violations; it is also a vivid illustration of a language game, showcasing how language strategic changes in the game between users and attackers. Investigating language game in open social media contexts is crucial for safeguarding personal information, enhancing social media governance, and fostering advancements in natural language processing research. This paper delves into the language game users play in open social media environments, employing a multi-agent simulation grounded in Large Language Models (LLMs). The agents embody two roles: the malicious agent, which infers privacy information, and the participant agents, who alter their communication styles when posting, thereby simulating language strategy dynamics under malicious monitoring to avoid possible privacy disclosure. The study assesses the framework’s efficacy through three simulation scenarios. The key findings suggest that LLMs can adeptly simulate the nuanced interplay of language and interactions against malicious monitoring, achieving a balance between privacy preservation and information fidelity. Additionally, it was observed that LLM-based agents employ distinct language game strategies tailored to specific scenarios.