Unsupervised opinion summarization aims to generate concise summaries which capture vital opinions from online reviews without any ground truth labels. However, most approaches suffer from the hallucination problem, generating inaccurate content. To tackle this problem, recent approaches focus on how to leverage domain-specific metadata. However, to employ these specific metadata, such approaches are based on some specific model structures, lacking generalization. Moreover, the effectiveness of these approaches is dependent on the availability of these domain metadata and lack of flexibility. Therefore, inspired by adversarial learning, we propose a model-agnostic and metadata-free adversarial framework (M2A) for unsupervised opinion summarization. In specific, natural language inference is appended to the generation model as the discriminator regardless of the structure of the generation model. Moreover, to avoid employing domain-specific metadata, the discriminator is retrained to align with the current domain through unsupervised contrastive learning. In this way, the discriminator is guaranteed metadata-free while obtaining general supervision. To show model-agnostic, we apply M2A to two different unsupervised summary generation models. Experimental results on the Amazon dataset show that M2A can generate comparable results on the ROUGE scores. Moreover, it significantly outperforms some state-of-the-art baselines on category accuracy and sentiment accuracy when evaluating summarization faithfulness.