Motivated by the subversion of “trusted” public parameters in mass-surveillance activities, we study the security of non-interactive zero-knowledge (NIZK) proofs in the presence of a maliciously chosen common reference string. We first provide definitions for subversion-resistant soundness, witness indistinguishability and zero knowledge. We show that certain combinations of goals are unachievable but for all other combinations we give constructions that achieve them. We then turn to zk-SNARKs (succinct non-interactive arguments of knowledge), which are computationally sound NIZK systems with short and efficiently verifiable proofs, used e.g. in cryptocurrencies such as Zcash. We show that under plausible hardness assumptions, many zk-SNARK schemes proposed in the literature can be made subversion-zero-knowledge at very little cost. (joint work with Mihir Bellare and Alessandra Scafuro)