Semantic communication, due to its focus on the transmitting meaning rather than the raw bit data, poses unique security challenges compared to the traditional communication systems. In particular, semantic communication systems are vulnerable to the malicious attacks that focus on the semantic layer, with the intention of understanding or distorting the intended meaning of the transmitted privacy data. Diffusion models, a class of generative artificial intelligence (GenAI), are well-suited for ensuring data security to attack. Through iteratively adding and then removing noise, diffusion models can generate meaningful information despite the presence of the unknown noise. This article proposes a diffusion-based framework to enhance the security of semantic transmission for the attacks including eavesdropping and jamming. Specifically, the proposed framework incorporates both the artificial noise and natural channel noise into the forward process of the diffusion models during the semantic transmission, with the reverse process used to remove noise at the legitimate receiver. In the eavesdropping scenarios, the artificial noise is the friendly noise designed to prevent semantic eavesdropping. In the jamming scenarios, the artificial noise is the malicious jamming generated by the jammer, which disrupts the semantic transmission. The case studies show that the proposed diffusion-based framework is promising in securing the semantic transmission. We also consolidate several broad research directions associated with the proposed framework.