Explainable recommender systems (XRSs) enhance user trust by providing personalized recommendations followed by persuasive explanations. Integrating large language models (LLMs), such as GPT-4, advances this domain but introduces risks from biases embedded within LLMs. These biases can lead XRSs to generate persuasive explanations that promote favored recommendations, influencing users to accept the model's preferences over their own. This paper identifies a previously unrecognized security threat: the intentional induction of XRSs via biased LLMs to promote specific items through misleading yet compelling explanations. Inspired by work in the psychology of persuasion, we construct biased datasets and systematically insert these biases into LLM-based XRSs. Experiments across four leading LLMs reveal that biases can significantly affect user decisions, with close to 50\% of users changing their choices. To counteract this, we propose a prompt rephrasing defense that effectively mitigates these biases, safeguarding the trustworthiness of XRSs.