We revisit the definitions of zero-knowledge in the Common Reference String (CRS) model and the Random Oracle (RO) model. We argue that even though these definitions syntactically mimic the standard zero-knowledge definition, they loose some of its spirit. In particular, we show that there exist a specific natural security property that is not captured by these definitions. This is the property of deniability. We formally define the notion of deniable zero-knowledge in these models and investigate the possibility of achieving it. Our results are different for the two models: – Concerning the CRS model, we rule out the possibility of achieving deniable zero-knowledge protocols in “natural” settings where such protocols cannot already be achieved in plain model. – In the RO model, on the other hand, we construct an efficient 2round deniable zero-knowledge argument of knowledge, that preserves both the zero-knowledge property and the proof of knowledge property under concur...